Contents

SciKits Numpy Matplotlib I Getting started with Python for science 2

1 Python scientific computing ecosystem 4
1.1 Why Python? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2025
1.2 The scientific Python ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Before starting: Installing a working environment . . . . . . . . . . . . . . . . . . . . . . 7
1.4 The workflow: interactive environments and text editors . . . . . . . . . . . . . . . . . . 7
SciPy Python EDITION
2 The Python language 12
2.1 First steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Basic types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Control Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Defining functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
IP[y]: 2.5 Reusing code: scripts and modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.6 Input and Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

Cython IPython
2.7 Standard Library . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
2.8 Exception handling in Python . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.9 Object-oriented programming (OOP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

3 NumPy: creating and manipulating numerical data 54

Edited by 3.1 The NumPy array object . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

Scien�fic Python
3.2 Numerical operations on arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
K. Jarrod Millman 3.3 More elaborate arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Stéfan van der Walt 3.4 Advanced operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Gaël Varoquaux 3.5 Some exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

Lectures
Emmanuelle Gouillart 3.6 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Olaf Vahtras
Pierre de Buyl 4 Matplotlib: plotting 108
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.2 Simple plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
lectures.scientific-python.org 4.3 Figures, Subplots, Axes and Ticks .
4.4 Other Types of Plots: examples and
. . . . . .
exercises .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
118
121
4.5 Beyond this tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4.6 Quick references . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4.7 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Gaël Varoquaux • Emmanuelle Gouillart • Olav Vahtras
Pierre de Buyl • K. Jarrod Millman • Stéfan van der Walt 5 SciPy : high-level scientific computing 211
5.1 File input/output: scipy.io . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
Christopher Burns • Adrian Chauve • Robert Cimrman • Christophe Combelles 5.2 Special functions: scipy.special . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
Ralf Gommers • André Espaze • Zbigniew Jędrzejewski-Szmek
Valen� n Haenel • Michael Hartmann • Gert-Ludwig Ingold • Fabian Pedregosa
i
Didrik Pinte • Nicolas P. Rougier • Joris Van den Bossche • Pauli Virtanen
and many others...
 5.3 Linear algebra operations: scipy.linalg . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 13.3 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
5.4 Interpolation: scipy.interpolate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 13.4 Examples for the mathematical optimization chapter . . . . . . . . . . . . . . . . . . . . 442
5.5 Optimization and fit: scipy.optimize . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218 13.5 Practical guide to optimization with SciPy . . . . . . . . . . . . . . . . . . . . . . . . . . 480
5.6 Statistics and random numbers: scipy.stats . . . . . . . . . . . . . . . . . . . . . . . . 223 13.6 Special case: non-linear least-squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 482
5.7 Numerical integration: scipy.integrate . . . . . . . . . . . . . . . . . . . . . . . . . . . 224 13.7 Optimization with constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483
5.8 Fast Fourier transforms: scipy.fft . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 13.8 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
5.9 Signal processing: scipy.signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229 13.9 Examples for the mathematical optimization chapter . . . . . . . . . . . . . . . . . . . . 485
5.10 Image manipulation: scipy.ndimage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
5.11 Summary exercises on scientific computing . . . . . . . . . . . . . . . . . . . . . . . . . . 237 14 Interfacing with C 486
5.12 Full code examples for the SciPy chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . 250 14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487
14.2 Python-C-Api . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487
6 Getting help and finding documentation 291 14.3 Ctypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495
14.4 SWIG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
14.5 Cython . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 504
II Advanced topics 294 14.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
14.7 Further Reading and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
7 Advanced Python Constructs 296 14.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
7.1 Iterators, generator expressions and generators . . . . . . . . . . . . . . . . . . . . . . . . 297
7.2 Decorators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
7.3 Context managers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310 III Packages and applications 511
8 Advanced NumPy 313 15 Statistics in Python 513
8.1 Life of ndarray . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314 15.1 Data representation and interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514
8.2 Universal functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328 15.2 Hypothesis testing: comparing two groups . . . . . . . . . . . . . . . . . . . . . . . . . . 519
8.3 Interoperability features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337 15.3 Linear models, multiple factors, and analysis of variance . . . . . . . . . . . . . . . . . . . 522
8.4 Array siblings: chararray, maskedarray . . . . . . . . . . . . . . . . . . . . . . . . . . . 340 15.4 More visualization: seaborn for statistical exploration . . . . . . . . . . . . . . . . . . . . 528
8.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343 15.5 Testing for interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 532
8.6 Contributing to NumPy/SciPy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343 15.6 Full code for the figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 533
15.7 Solutions to this chapter’s exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 561
9 Debugging code 346
9.1 Avoiding bugs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347 16 Sympy : Symbolic Mathematics in Python 564
9.2 Debugging workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349 16.1 First Steps with SymPy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565
9.3 Using the Python debugger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349 16.2 Algebraic manipulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 566
9.4 Debugging segmentation faults using gdb . . . . . . . . . . . . . . . . . . . . . . . . . . . 355 16.3 Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
16.4 Equation solving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
10 Optimizing code 358 16.5 Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 570
10.1 Optimization workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
10.2 Profiling Python code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359 17 scikit-image: image processing 572
10.3 Making code go faster . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362 17.1 Introduction and concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573
10.4 Writing faster numerical code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363 17.2 Importing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 575
17.3 Example data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 576
11 Sparse Arrays in SciPy 366 17.4 Input/output, data types and colorspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 576
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366 17.5 Image preprocessing / enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 578
11.2 Storage Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368 17.6 Image segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581
11.3 Linear System Solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381 17.7 Measuring regions’ properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 584
11.4 Other Interesting Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386 17.8 Data visualization and interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585
17.9 Feature extraction for computer vision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585
12 Image manipulation and processing using NumPy and SciPy 387
17.10 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 586
12.1 Opening and writing to image files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388
17.11 Examples for the scikit-image chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 586
12.2 Displaying images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390
12.3 Basic manipulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 391 18 scikit-learn: machine learning in Python 598
12.4 Image filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393 18.1 Introduction: problem settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
12.5 Feature extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398 18.2 Basic principles of machine learning with scikit-learn . . . . . . . . . . . . . . . . . . . . 603
12.6 Measuring objects properties: scipy.ndimage.measurements . . . . . . . . . . . . . . . . 401 18.3 Supervised Learning: Classification of Handwritten Digits . . . . . . . . . . . . . . . . . . 608
12.7 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405 18.4 Supervised Learning: Regression of Housing Data . . . . . . . . . . . . . . . . . . . . . . 613
12.8 Examples for the image processing chapter . . . . . . . . . . . . . . . . . . . . . . . . . . 405 18.5 Measuring prediction performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 618
18.6 Unsupervised Learning: Dimensionality Reduction and Visualization . . . . . . . . . . . . 623
13 Mathematical optimization: finding minima of functions 432
18.7 Parameter selection, Validation, and Testing . . . . . . . . . . . . . . . . . . . . . . . . . 626
13.1 Knowing your problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433
18.8 Examples for the scikit-learn chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 634
13.2 A review of the different optimizers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435

ii iii
 Scientific Python Lectures, Edition 2025.1rc0.dev0

IV About the Scientific Python Lectures 681

19 About the Scientific Python Lectures 682
19.1 Authors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 682

Index 685

iv Contents 1
 Scientific Python Lectures, Edition 2025.1rc0.dev0

This part of the Scientific Python Lectures is a self-contained introduction to everything that is needed
to use Python for science, from the language itself, to numerical computing or plotting.

Part I

Getting started with Python for

science

2 3
 Scientific Python Lectures, Edition 2025.1rc0.dev0

• Universal Python is a language used for many different problems. Learning Python avoids learning
a new software for each new problem.

1.1.3 How does Python compare to other solutions?

1
Compiled languages: C, C++, Fortran. . .
Pros
• Very fast. For heavy computations, it’s difficult to outperform these languages.
Cons

CHAPTER • Painful usage: no interactivity during development, mandatory compilation steps,

verbose syntax, manual memory management. These are difficult languages for
non programmers.

Matlab scripting language

Pros
• Very rich collection of libraries with numerous algorithms, for many different do-
mains. Fast execution because these libraries are often written in a compiled lan-

Python scientific computing ecosystem guage.

• Pleasant development environment: comprehensive and help, integrated editor, etc.
• Commercial support is available.
Cons
• Base language is quite poor and can become restrictive for advanced users.
• Not free and not everything is open sourced.
Authors: Fernando Perez, Emmanuelle Gouillart, Gaël Varoquaux, Valentin Haenel
Julia

1.1 Why Python? Pros

• Fast code, yet interactive and simple.
1.1.1 The scientist’s needs
• Easily connects to Python or C.
• Get data (simulation, experiment control),
Cons
• Manipulate and process data,
• Ecosystem limited to numerical computing.
• Visualize results, quickly to understand, but also with high quality figures, for reports or publica-
• Still young.
tions.

Other scripting languages: Scilab, Octave, R, IDL, etc.

1.1.2 Python’s strengths
Pros
• Batteries included Rich collection of already existing bricks of classic numerical methods, plot-
ting or data processing tools. We don’t want to re-program the plotting of a curve, a Fourier • Open-source, free, or at least cheaper than Matlab.
transform or a fitting algorithm. Don’t reinvent the wheel! • Some features can be very advanced (statistics in R, etc.)
• Easy to learn Most scientists are not paid as programmers, neither have they been trained so. Cons
They need to be able to draw a curve, smooth a signal, do a Fourier transform in a few minutes.
• Fewer available algorithms than in Matlab, and the language is not more advanced.
• Easy communication To keep code alive within a lab or a company it should be as readable
as a book by collaborators, students, or maybe customers. Python syntax is simple, avoiding • Some software are dedicated to one domain. Ex: Gnuplot to draw curves. These
strange symbols or lengthy routine specifications that would divert the reader from mathematical programs are very powerful, but they are restricted to a single type of usage, such
or scientific understanding of the code. as plotting.
• Efficient code Python numerical modules are computationally efficient. But needless to say that
a very fast code becomes useless if too much time is spent writing it. Python aims for quick
development times and quick execution times.

4 1.1. Why Python? 5

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Python • Matplotlib : 2-D visualization, “publication-ready” plots https://matplotlib.org/

Pros
• Very rich scientific computing libraries ã See also

• Well thought out language, allowing to write very readable and well structured code: chapter on matplotlib
we “code what we think”.
• Many libraries beyond scientific computing (web server, serial port access, etc.) Advanced interactive environments:
• Free and open-source software, widely spread, with a vibrant community. • IPython, an advanced Python console https://ipython.org/
• A variety of powerful environments to work in, such as IPython, Spyder, Jupyter • Jupyter, notebooks in the browser https://jupyter.org/
notebooks, Pycharm, Visual Studio Code
Domain-specific packages,
Cons
• pandas, statsmodels, seaborn for statistics
• Not all the algorithms that can be found in more specialized software or toolboxes.
• sympy for symbolic computing
• scikit-image for image processing
1.2 The scientific Python ecosystem
• scikit-learn for machine learning
Unlike Matlab, or R, Python does not come with a pre-bundled set of modules for scientific computing. and many more packages not documented in the Scientific Python Lectures.
Below are the basic building blocks that can be combined to obtain a scientific computing environment:

ã See also

chapters on advanced topics

chapters on packages and applications
Python, a generic and modern computing language
• The language: flow control, data types (string, int), data collections (lists, dictionaries), etc.
• Modules of the standard library: string processing, file management, simple network protocols.
• A large number of specialized modules or applications written in Python: web framework, etc. . . . 1.3 Before starting: Installing a working environment
and scientific computing.
Python comes in many flavors, and there are many ways to install it. However, we recommend to install
• Development tools (automatic testing, documentation generation) a scientific-computing distribution, that comes readily with optimized versions of scientific modules.
Under Linux
ã See also
If you have a recent distribution, most of the tools are probably packaged, and it is recommended to use
chapter on Python language your package manager.
Other systems
Core numeric libraries
There are several fully-featured scientific Python distributions:
• NumPy: numerical computing with powerful numerical arrays objects, and routines to manip-
• Anaconda
ulate them. https://numpy.org/
• WinPython

ã See also
1.4 The workflow: interactive environments and text editors
chapter on numpy
Interactive work to test and understand algorithms: In this section, we describe a workflow
• SciPy : high-level numerical routines. Optimization, regression, interpolation, etc https://scipy. combining interactive work and consolidation.
org/ Python is a general-purpose language. As such, there is not one blessed environment to work in, and
not only one way of using it. Although this makes it harder for beginners to find their way, it makes it
possible for Python to be used for programs, in web servers, or embedded devices.
ã See also

chapter on SciPy

1.2. The scientific Python ecosystem 6 1.3. Before starting: Installing a working environment 7
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

1.4.1 Interactive work In [3]: %run my_file.py

We recommend an interactive work with the IPython console, or its offspring, the Jupyter notebook. Hello world
They are handy to explore and understand algorithms.
In [4]: s
Out[4]: 'Hello world'
Under the notebook
In [5]: %whos
To execute code, press “shift enter”
Variable Type Data/Info
----------------------------
Start ipython: s str Hello world
In [1]: print('Hello world')
Hello world
From a script to functions
Getting help by using the ? operator after an object:
While it is tempting to work only with scripts, that is a file full of instructions following each other,
In [2]: print? do plan to progressively evolve the script to a set of functions:
Signature: print(*args, sep=' ', end='\n', file=None, flush=False) • A script is not reusable, functions are.
Docstring:
Prints the values to a stream, or to sys.stdout by default. • Thinking in terms of functions helps breaking the problem in small blocks.

sep
string inserted between values, default a space. 1.4.3 IPython and Jupyter Tips and Tricks
end The user manuals contain a wealth of information. Here we give a quick introduction to four useful
string appended after the last value, default a newline. features: history, tab completion, magic functions, and aliases.
file
a file-like object (stream); defaults to the current sys.stdout.
flush
whether to forcibly flush the stream.
Type: builtin_function_or_method
Command history Like a UNIX shell, the IPython console supports command history. Type up and
down to navigate previously typed commands:
ã See also
In [6]: x = 10
• IPython user manual: https://ipython.readthedocs.io/en/stable/
In [7]: <UP>
• Jupyter Notebook QuickStart: https://docs.jupyter.org/en/latest/start/index.html
In [8]: x = 10
1.4.2 Elaboration of the work in an editor
As you move forward, it will be important to not only work interactively, but also to create and reuse
Python files. For this, a powerful code editor will get you far. Here are several good easy-to-use editors:
• Spyder: integrates an IPython console, a debugger, a profiler. . .
Tab completion Tab completion, is a convenient way to explore the structure of any object you’re
• PyCharm: integrates an IPython console, notebooks, a debugger. . . (freely available, but commer-
dealing with. Simply type object_name.<TAB> to view the object’s attributes. Besides Python objects
cial)
and keywords, tab completion also works on file and directory names.*
• Visual Studio Code: integrates a Python console, notebooks, a debugger, . . .
In [9]: x = 10
Some of these are shipped by the various scientific Python distributions, and you can find them in the
menus. In [10]: x.<TAB>
As an exercise, create a file my_file.py in a code editor, and add the following lines: as_integer_ratio() conjugate() imag to_bytes()
bit_count() denominator numerator
s = 'Hello world' bit_length() from_bytes() real
print(s)

Now, you can run it in IPython console or a notebook and explore the resulting variables:

1.4. The workflow: interactive environments and text editors 8 1.4. The workflow: interactive environments and text editors 9
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Magic functions The console and the notebooks support so-called magic functions by prefixing a
Chapter on debugging
command with the % character. For example, the run and whos functions from the previous section are
magic functions. Note that, the setting automagic, which is enabled by default, allows you to omit the
preceding % sign. Thus, you can just type the magic function and it will work.
Other useful magic functions are:
• %cd to change the current directory.
Aliases Furthermore IPython ships with various aliases which emulate common UNIX command line
In [11]: cd /tmp
tools such as ls to list files, cp to copy files and rm to remove files (a full list of aliases is shown when
/tmp
typing alias).
• %cpaste allows you to paste code, especially code from websites which has been prefixed with the
standard Python prompt (e.g. >>>) or with an ipython prompt, (e.g. in [3]): Getting help
In [12]: %cpaste • The built-in cheat-sheet is accessible via the %quickref magic function.
• A list of all available magic functions is shown when typing %magic.
• %timeit allows you to time the execution of short snippets using the timeit module from the
standard library:

In [12]: %timeit x = 10
11.8 ns +- 0.0116 ns per loop (mean +- std. dev. of 7 runs, 100,000,000 loops␣
˓→each)

ã See also

Chapter on optimizing code

• %debug allows you to enter post-mortem debugging. That is to say, if the code you try to execute,
raises an exception, using %debug will enter the debugger at the point where the exception was
thrown.

In [13]: x === 10
Cell In[13], line 1
x === 10
^
SyntaxError: invalid syntax

In [14]: %debug
> /home/jarrod/.venv/lectures/lib64/python3.11/site-packages/IPython/core/
˓→compilerop.py(86)ast_parse()

84 Arguments are exactly the same as ast.parse (in the standard␣

˓→library),

85 and are passed to the built-in compile function."""

---> 86 return compile(source, filename, symbol, self.flags | PyCF_ONLY_
˓→AST, 1)

87
88 def reset_compiler_flags(self):
ipdb> locals()
{'self': <IPython.core.compilerop.CachingCompiler object at 0x7f30d02efc10>,
˓→'source': 'x === 10\n', 'filename': '<ipython-input-1-8e8bc565444b>', 'symbol

˓→': 'exec'}

ipdb>

ã See also

1.4. The workflow: interactive environments and text editors 10 1.4. The workflow: interactive environments and text editors 11
 Scientific Python Lectures, Edition 2025.1rc0.dev0

• a very readable language with clear non-verbose syntax

• a language for which a large variety of high-quality packages are available for various applica-
tions, from web frameworks to scientific computing.
• a language very easy to interface with other languages, in particular C and C++.

2
• Some other features of the language are illustrated just below. For example, Python is an object-
oriented language, with dynamic typing (the same variable can contain objects of different types
during the course of a program).
See https://www.python.org/about/ for more information about distinguishing features of Python.

CHAPTER

2.1 First steps

Start the Ipython shell (an enhanced interactive Python shell):
• by typing “ipython” from a Linux/Mac terminal, or from the Windows cmd shell,

The Python language

• or by starting the program from a menu, e.g. the Anaconda Navigator, the Python(x,y) menu if
you have installed one of these scientific-Python suites.

 Tip

If you don’t have Ipython installed on your computer, other Python shells are available, such as the
plain Python shell started by typing “python” in a terminal, or the Idle interpreter. However, we
advise to use the Ipython shell because of its enhanced features, especially for interactive scientific
computing.
Authors: Chris Burns, Christophe Combelles, Emmanuelle Gouillart, Gaël Varoquaux

Once you have started the interpreter, type

Python for scientific computing
>>> print("Hello, world!")
We introduce here the Python language. Only the bare minimum necessary for getting started with
Hello, world!
NumPy and SciPy is addressed here. To learn more about the language, consider going through
the excellent tutorial https://docs.python.org/3/tutorial. Dedicated books are also available, such as
Dive into Python 3.
 Tip

The message “Hello, world!” is then displayed. You just executed your first Python instruction,
congratulations!

To get yourself started, type the following stack of instructions

>>> a = 3
 Tip >>> b = 2*a
Python is a programming language, as are C, Fortran, BASIC, PHP, etc. Some specific features >>> type(b)
of Python are as follows: <class 'int'>
>>> print(b)
• an interpreted (as opposed to compiled) language. Contrary to e.g. C or Fortran, one does 6
not compile Python code before executing it. In addition, Python can be used interactively: >>> a*b
many Python interpreters are available, from which commands and scripts can be executed. 18
• a free software released under an open-source license: Python can be used and distributed >>> b = 'hello'
free of charge, even for building commercial software. >>> type(b)
<class 'str'>
• multi-platform: Python is available for all major operating systems, Windows, Linux/Unix, >>> b + b
MacOS X, most likely your mobile phone OS, etc. 'hellohello'
(continues on next page)

12 2.1. First steps 13

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

>>> 2*b >>> test = (3 > 4)
'hellohello' >>> test
False
>>> type(test)
 Tip <class 'bool'>

Two variables a and b have been defined above. Note that one does not declare the type of a variable
before assigning its value. In C, conversely, one should write:  Tip
int a = 3;
A Python shell can therefore replace your pocket calculator, with the basic arithmetic operations +,
In addition, the type of a variable may change, in the sense that at one point in time it can be equal -, *, /, % (modulo) natively implemented
to a value of a certain type, and a second point in time, it can be equal to a value of a different type.
b was first equal to an integer, but it became equal to a string when it was assigned the value ‘hello’.
Operations on integers (b=2*a) are coded natively in Python, and so are some operations on strings >>> 7 * 3.
such as additions and multiplications, which amount respectively to concatenation and repetition. 21.0
>>> 2**10
1024
>>> 8 % 3
2

2.2 Basic types Type conversion (casting):

2.2.1 Numerical types >>> float(1)

1.0

 Tip
2.2.2 Containers
Python supports the following numerical, scalar types:

 Tip
Integer
Python provides many efficient types of containers, in which collections of objects can be stored.
>>> 1 + 1
2
>>> a = 4 Lists
>>> type(a)
<class 'int'>
 Tip
Floats
A list is an ordered collection of objects, that may have different types. For example:
>>> c = 2.1
>>> type(c)
>>> colors = ['red', 'blue', 'green', 'black', 'white']
<class 'float'>
>>> type(colors)
<class 'list'>
Complex

>>> a = 1.5 + 0.5j Indexing: accessing individual objects contained in the list:
>>> a.real
>>> colors[2]
1.5
'green'
>>> a.imag
0.5
>>> type(1. + 0j) Counting from the end with negative indices:
<class 'complex'> >>> colors[-1]
'white'
Booleans >>> colors[-2]
'black'
>>> 3 > 4
False
(continues on next page)

2.2. Basic types 14 2.2. Basic types 15

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

. Warning
 Tip
Indexing starts at 0 (as in C), not at 1 (as in Fortran or Matlab)! For collections of numerical data that all have the same type, it is often more efficient to use
the array type provided by the numpy module. A NumPy array is a chunk of memory containing
Slicing: obtaining sublists of regularly-spaced elements: fixed-sized items. With NumPy arrays, operations on elements can be faster because elements are
regularly spaced in memory and more operations are performed through specialized C functions
>>> colors instead of Python loops.
['red', 'blue', 'green', 'black', 'white']
>>> colors[2:4]
['green', 'black']
 Tip

. Warning Python offers a large panel of functions to modify lists, or query them. Here are a few examples; for
more details, see https://docs.python.org/3/tutorial/datastructures.html#more-on-lists
Note that colors[start:stop] contains the elements with indices i such as start<= i < stop (i
ranging from start to stop-1). Therefore, colors[start:stop] has (stop - start) elements.
Add and remove elements:

Slicing syntax: colors[start:stop:stride] >>> colors = ['red', 'blue', 'green', 'black', 'white']
>>> colors.append('pink')
>>> colors
 Tip ['red', 'blue', 'green', 'black', 'white', 'pink']
>>> colors.pop() # removes and returns the last item
All slicing parameters are optional: 'pink'
>>> colors >>> colors
['red', 'blue', 'green', 'black', 'white'] ['red', 'blue', 'green', 'black', 'white']
>>> colors[3:] >>> colors.extend(['pink', 'purple']) # extend colors, in-place
['black', 'white'] >>> colors
>>> colors[:3] ['red', 'blue', 'green', 'black', 'white', 'pink', 'purple']
['red', 'blue', 'green'] >>> colors = colors[:-2]
>>> colors[::2] >>> colors
['red', 'green', 'white'] ['red', 'blue', 'green', 'black', 'white']

Reverse:
Lists are mutable objects and can be modified:
>>> rcolors = colors[::-1]
>>> colors[0] = 'yellow' >>> rcolors
>>> colors ['white', 'black', 'green', 'blue', 'red']
['yellow', 'blue', 'green', 'black', 'white'] >>> rcolors2 = list(colors) # new object that is a copy of colors in a different␣
>>> colors[2:4] = ['gray', 'purple'] ˓→memory area

>>> colors >>> rcolors2

['yellow', 'blue', 'gray', 'purple', 'white'] ['red', 'blue', 'green', 'black', 'white']
>>> rcolors2.reverse() # in-place; reversing rcolors2 does not affect colors
>>> rcolors2
['white', 'black', 'green', 'blue', 'red']
ò Note

The elements of a list may have different types: Concatenate and repeat lists:
>>> colors = [3, -200, 'hello'] >>> rcolors + colors
>>> colors ['white', 'black', 'green', 'blue', 'red', 'red', 'blue', 'green', 'black', 'white']
[3, -200, 'hello'] >>> rcolors * 2
>>> colors[1], colors[2] ['white', 'black', 'green', 'blue', 'red', 'white', 'black', 'green', 'blue', 'red']
(-200, 'hello')

 Tip

Sort:

2.2. Basic types 16 2.2. Basic types 17

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

>>> sorted(rcolors) # new object >>> a = "hello"

['black', 'blue', 'green', 'red', 'white'] >>> a[0]
>>> rcolors 'h'
['white', 'black', 'green', 'blue', 'red'] >>> a[1]
>>> rcolors.sort() # in-place 'e'
>>> rcolors >>> a[-1]
['black', 'blue', 'green', 'red', 'white'] 'o'

 Tip
Methods and Object-Oriented Programming
(Remember that negative indices correspond to counting from the right end.)
The notation rcolors.method() (e.g. rcolors.append(3) and colors.pop()) is our first example
of object-oriented programming (OOP). Being a list, the object rcolors owns the method function
that is called using the notation .. No further knowledge of OOP than understanding the notation . Slicing:
is necessary for going through this tutorial.
>>> a = "hello, world!"
>>> a[3:6] # 3rd to 6th (excluded) elements: elements 3, 4, 5
'lo,'
Discovering methods:
>>> a[2:10:2] # Syntax: a[start:stop:step]
Reminder: in Ipython: tab-completion (press tab) 'lo o'
>>> a[::3] # every three characters, from beginning to end
In [1]: rcolors.<TAB> 'hl r!'
append() count() insert() reverse()
clear() extend() pop() sort()
copy() index() remove()
 Tip

Accents and special characters can also be handled as in Python 3 strings consist of Unicode charac-
Strings ters.
Different string syntaxes (simple, double or triple quotes):
A string is an immutable object and it is not possible to modify its contents. One may however create
s = 'Hello, how are you?' new strings from the original one.
s = "Hi, what's up"
s = '''Hello, In [3]: a = "hello, world!"
how are you''' # tripling the quotes allows the
# string to span more than one line In [4]: a.replace('l', 'z', 1)
s = """Hi, Out[4]: 'hezlo, world!'
what's up?"""

In [2]: 'Hi, what's up?'  Tip

Cell In[2], line 1
Strings have many useful methods, such as a.replace as seen above. Remember the a. object-
'Hi, what's up?'
oriented notation and use tab completion or help(str) to search for new methods.
^
SyntaxError: unterminated string literal (detected at line 1)

This syntax error can be avoided by enclosing the string in double quotes instead of single quotes. ã See also
Alternatively, one can prepend a backslash to the second single quote. Other uses of the backslash are,
e.g., the newline character \n and the tab character \t. Python offers advanced possibilities for manipulating strings, looking for patterns or formatting. The
interested reader is referred to https://docs.python.org/3/library/stdtypes.html#string-methods and
https://docs.python.org/3/library/string.html#format-string-syntax
 Tip

Strings are collections like lists. Hence they can be indexed and sliced, using the same syntax and String formatting:
rules.
>>> 'An integer: %i ; a float: %f ; another string: %s ' % (1, 0.1, 'string') # with␣
˓→more values use tuple after %
Indexing:
'An integer: 1; a float: 0.100000; another string: string'
(continues on next page)

2.2. Basic types 18 2.2. Basic types 19

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> s = set(('a', 'b', 'c', 'a'))
>>> s
>>> i = 102 {'a', 'b', 'c'}
>>> filename = 'processing_of_dataset_%d .txt' % i # no need for tuples with just␣ >>> s.difference(('a', 'b'))
˓→one value after %
{'c'}
>>> filename
'processing_of_dataset_102.txt'
2.2.3 Assignment operator
Dictionaries
 Tip
 Tip Python library reference says:
A dictionary is basically an efficient table that maps keys to values. Assignment statements are used to (re)bind names to values and to modify attributes or
items of mutable objects.
In short, it works as follows (simple assignment):
>>> tel = {'emmanuelle': 5752, 'sebastian': 5578}
>>> tel['francis'] = 5915 1. an expression on the right hand side is evaluated, the corresponding object is created/obtained
>>> tel
2. a name on the left hand side is assigned, or bound, to the r.h.s. object
{'emmanuelle': 5752, 'sebastian': 5578, 'francis': 5915}
>>> tel['sebastian']
5578 Things to note:
>>> tel.keys()
dict_keys(['emmanuelle', 'sebastian', 'francis']) • A single object can have several names bound to it:
>>> tel.values() In [5]: a = [1, 2, 3]
dict_values([5752, 5578, 5915])
>>> 'francis' in tel In [6]: b = a
True
In [7]: a
Out[7]: [1, 2, 3]
 Tip
In [8]: b
It can be used to conveniently store and retrieve values associated with a name (a string for a date, Out[8]: [1, 2, 3]
a name, etc.). See https://docs.python.org/3/tutorial/datastructures.html#dictionaries for more
information. In [9]: a is b
A dictionary can have keys (resp. values) with different types: Out[9]: True

>>> d = {'a':1, 'b':2, 3:'hello'} In [10]: b[1] = 'hi!'

>>> d
{'a': 1, 'b': 2, 3: 'hello'} In [11]: a
Out[11]: [1, 'hi!', 3]

More container types • to change a list in place, use indexing/slices:

Tuples In [12]: a = [1, 2, 3]
Tuples are basically immutable lists. The elements of a tuple are written between parentheses, or just
separated by commas: In [13]: a
Out[13]: [1, 2, 3]
>>> t = 12345, 54321, 'hello!'
>>> t[0] In [14]: a = ['a', 'b', 'c'] # Creates another object.
12345
>>> t In [15]: a
(12345, 54321, 'hello!') Out[15]: ['a', 'b', 'c']
>>> u = (0, 2)
In [16]: id(a)
Sets: unordered, unique items: Out[16]: 139794116546752

(continues on next page)

2.2. Basic types 20 2.2. Basic types 21

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) 2.3.2 for/range

In [17]: a[:] = [1, 2, 3] # Modifies object in place. Iterating with an index:
In [18]: a >>> for i in range(4):
Out[18]: [1, 2, 3] ... print(i)
0
In [19]: id(a) 1
Out[19]: 139794116546752 2
3
• the key concept here is mutable vs. immutable
But most often, it is more readable to iterate over values:
– mutable objects can be changed in place
– immutable objects cannot be modified once created >>> for word in ('cool', 'powerful', 'readable'):
... print('Python is %s ' % word)
Python is cool
ã See also Python is powerful
Python is readable
A very good and detailed explanation of the above issues can be found in David M. Beazley’s article
Types and Objects in Python.
2.3.3 while/break/continue
Typical C-style while loop (Mandelbrot problem):

>>> z = 1 + 1j
2.3 Control Flow >>> while abs(z) < 100:
... z = z**2 + 1
Controls the order in which the code is executed. >>> z
(-134+352j)
2.3.1 if/elif/else
More advanced features
>>> if 2**2 == 4:
... print("Obvious!") break out of enclosing for/while loop:
...
>>> z = 1 + 1j
Obvious!
>>> while abs(z) < 100:
Blocks are delimited by indentation
... if z.imag == 0:
... break
 Tip ... z = z**2 + 1

Type the following lines in your Python interpreter, and be careful to respect the indentation continue the next iteration of a loop.:
depth. The Ipython shell automatically increases the indentation depth after a colon : sign; to
decrease the indentation depth, go four spaces to the left with the Backspace key. Press the Enter >>> a = [1, 0, 2, 4]
key twice to leave the logical block. >>> for element in a:
... if element == 0:
... continue
>>> a = 10 ... print(1. / element)
1.0
>>> if a == 1: 0.5
... print(1) 0.25
... elif a == 2:
... print(2)
... else: 2.3.4 Conditional Expressions
... print("A lot")
... if <OBJECT>
A lot Evaluates to False:
• any number equal to zero (0, 0.0, 0+0j)
Indentation is compulsory in scripts as well. As an exercise, re-type the previous lines with the same
indentation in a script condition.py, and execute the script with run condition.py in Ipython. • an empty container (list, tuple, set, dictionary, . . . )

2.3. Control Flow 22 2.3. Control Flow 23

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• False, None (continued from previous page)

Evaluates to True: how

are
• everything else you?
a == b
Tests equality, with logics:
 Tip
>>> 1 == 1.
True Few languages (in particular, languages for scientific computing) allow to loop over anything but
integers/indices. With Python it is possible to loop exactly over the objects of interest without
a is b bothering with indices you often don’t care about. This feature can often be used to make code more
Tests identity: both sides are the same object: readable.

>>> a = 1
>>> b = 1.
>>> a == b . Warning
True
>>> a is b Not safe to modify the sequence you are iterating over.
False
Keeping track of enumeration number
>>> a = 1
>>> b = 1 Common task is to iterate over a sequence while keeping track of the item number.
>>> a is b
• Could use while loop with a counter as above. Or a for loop:
True
>>> words = ('cool', 'powerful', 'readable')
a in b >>> for i in range(0, len(words)):
For any collection b: b contains a ... print((i, words[i]))
(0, 'cool')
>>> b = [1, 2, 3]
(1, 'powerful')
>>> 2 in b
(2, 'readable')
True
>>> 5 in b
False • But, Python provides a built-in function - enumerate - for this:

>>> for index, item in enumerate(words):

If b is a dictionary, this tests that a is a key of b. ... print((index, item))
(0, 'cool')
2.3.5 Advanced iteration (1, 'powerful')
(2, 'readable')
Iterate over any sequence
You can iterate over any sequence (string, list, keys in a dictionary, lines in a file, . . . ):
Looping over a dictionary
>>> vowels = 'aeiouy' Use items:
>>> for i in 'powerful': >>> d = {'a': 1, 'b':1.2, 'c':1j}
... if i in vowels:
... print(i) >>> for key, val in sorted(d.items()):
o ... print('Key: %s has value: %s ' % (key, val))
e Key: a has value: 1
u Key: b has value: 1.2
Key: c has value: 1j
>>> message = "Hello how are you?"
>>> message.split() # returns a list
['Hello', 'how', 'are', 'you?'] ò Note
>>> for word in message.split():
... print(word) The ordering of a dictionary is random, thus we use sorted() which will sort on the keys.
...
Hello
(continues on next page)

2.3. Control Flow 24 2.3. Control Flow 25

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

2.3.6 List Comprehensions

• the def keyword;
Instead of creating a list by means of a loop, one can make use of a list comprehension with a rather
• is followed by the function’s name, then
self-explaining syntax.
• the arguments of the function are given between parentheses followed by a colon.
>>> [i**2 for i in range(4)]
[0, 1, 4, 9] • the function body;
• and return object for optionally returning values.

Exercise 2.4.3 Parameters

Compute the decimals of Pi using the Wallis formula: Mandatory parameters (positional arguments)
∞
∏︁ 4𝑖2 In [5]: def double_it(x):
𝜋=2 ...: return x * 2
4𝑖2−1
𝑖=1 ...:

In [6]: double_it(3)
Out[6]: 6

2.4 Defining functions In [7]: double_it()

---------------------------------------------------------------------------
2.4.1 Function definition TypeError Traceback (most recent call last)
Cell In[7], line 1
In [1]: def test(): ----> 1 double_it()
...: print('in test function')
...: TypeError: double_it() missing 1 required positional argument: 'x'
...:
Optional parameters (keyword or named arguments)
In [2]: test()
in test function In [8]: def double_it(x=2):
...: return x * 2
...:
. Warning
In [9]: double_it()
Function blocks must be indented as other control-flow blocks. Out[9]: 4

In [10]: double_it(3)
2.4.2 Return statement Out[10]: 6
Functions can optionally return values.
Keyword arguments allow you to specify default values.
In [3]: def disk_area(radius):
...: return 3.14 * radius * radius
...: . Warning

Default values are evaluated when the function is defined, not when it is called. This can be prob-
In [4]: disk_area(1.5) lematic when using mutable types (e.g. dictionary or list) and modifying them in the function body,
Out[4]: 7.0649999999999995 since the modifications will be persistent across invocations of the function.
Using an immutable type in a keyword argument:
ò Note In [11]: bigx = 10

By default, functions return None. In [12]: def double_it(x=bigx):

....: return x * 2
....:
ò Note
In [13]: bigx = 1e9 # Now really big
Note the syntax to define a function:
In [14]: double_it()
Out[14]: 20

2.4. Defining functions 26 2.4. Defining functions 27

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

2.4.4 Passing by value

Using an mutable type in a keyword argument (and modifying it inside the function body):
In [15]: def add_to_dict(args={'a': 1, 'b': 2}):
....: for i in args.keys():  Tip
....: args[i] += 1
....: print(args) Can you modify the value of a variable inside a function? Most languages (C, Java, . . . ) distinguish
....: “passing by value” and “passing by reference”. In Python, such a distinction is somewhat artificial,
and it is a bit subtle whether your variables are going to be modified or not. Fortunately, there exist
In [16]: add_to_dict clear rules.
Out[16]: <function __main__.add_to_dict(args={'a': 1, 'b': 2})> Parameters to functions are references to objects, which are passed by value. When you pass a
variable to a function, python passes the reference to the object to which the variable refers (the
In [17]: add_to_dict() value). Not the variable itself.
{'a': 2, 'b': 3}

In [18]: add_to_dict() If the value passed in a function is immutable, the function does not modify the caller’s variable. If the
{'a': 3, 'b': 4} value is mutable, the function may modify the caller’s variable in-place:

In [19]: add_to_dict() >>> def try_to_modify(x, y, z):

{'a': 4, 'b': 5} ... x = 23
... y.append(42)
... z = [99] # new reference
... print(x)
 Tip ... print(y)
... print(z)
More involved example implementing python’s slicing: ...
>>> a = 77 # immutable variable
In [20]: def slicer(seq, start=None, stop=None, step=None):
>>> b = [99] # mutable variable
....: """Implement basic python slicing."""
>>> c = [28]
....: return seq[start:stop:step]
>>> try_to_modify(a, b, c)
....:
23
[99, 42]
In [21]: rhyme = 'one fish, two fish, red fish, blue fish'.split()
[99]
>>> print(a)
In [22]: rhyme
77
Out[22]: ['one', 'fish,', 'two', 'fish,', 'red', 'fish,', 'blue', 'fish']
>>> print(b)
[99, 42]
In [23]: slicer(rhyme)
>>> print(c)
Out[23]: ['one', 'fish,', 'two', 'fish,', 'red', 'fish,', 'blue', 'fish']
[28]
In [24]: slicer(rhyme, step=2)
Functions have a local variable table called a local namespace.
Out[24]: ['one', 'two', 'red', 'blue']
The variable x only exists within the function try_to_modify.
In [25]: slicer(rhyme, 1, step=2)
Out[25]: ['fish,', 'fish,', 'fish,', 'fish'] 2.4.5 Global variables
In [26]: slicer(rhyme, start=1, stop=4, step=2) Variables declared outside the function can be referenced within the function:
Out[26]: ['fish,', 'fish,']
In [28]: x = 5
The order of the keyword arguments does not matter:
In [29]: def addx(y):
In [27]: slicer(rhyme, step=2, start=1, stop=4) ....: return x + y
Out[27]: ['fish,', 'fish,'] ....:
but it is good practice to use the same ordering as the function’s definition.
In [30]: addx(10)
Out[30]: 15
Keyword arguments are a very convenient feature for defining functions with a variable number of argu-
ments, especially when default values are to be used in most calls to the function. But these “global” variables cannot be modified within the function, unless declared global in the
function.
This doesn’t work:

2.4. Defining functions 28 2.4. Defining functions 29

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

In [31]: def setx(y):
....: x = y In [40]: funcname?
....: print('x is %d ' % x) Signature: funcname(params)
....: Docstring:
....: Concise one-line sentence describing the function.
Extended summary which can contain multiple paragraphs.
In [32]: setx(10) File: ~/src/scientific-python-lectures/<ipython-input-13-64e466df6d64>
x is 10 Type: function

In [33]: x
Out[33]: 5 ò Note

This works: Docstring guidelines

For the sake of standardization, the Docstring Conventions webpage documents the semantics and
In [34]: def setx(y): conventions associated with Python docstrings.
....: global x
....: x = y Also, the NumPy and SciPy modules have defined a precise standard for documenting scientific func-
....: print('x is %d ' % x) tions, that you may want to follow for your own functions, with a Parameters section, an Examples
....: section, etc. See https://numpydoc.readthedocs.io/en/latest/format.html#docstring-standard
....:

In [35]: setx(10) 2.4.8 Functions are objects

x is 10
Functions are first-class objects, which means they can be:
In [36]: x • assigned to a variable
Out[36]: 10
• an item in a list (or any collection)
• passed as an argument to another function.
2.4.6 Variable number of parameters
In [41]: va = variable_args
Special forms of parameters:
• *args: any number of positional arguments packed into a tuple In [42]: va('three', x=1, y=2)
args is ('three',)
• **kwargs: any number of keyword arguments packed into a dictionary kwargs is {'x': 1, 'y': 2}
In [37]: def variable_args(*args, **kwargs):
....: print('args is', args)
2.4.9 Methods
....: print('kwargs is', kwargs)
....: Methods are functions attached to objects. You’ve seen these in our examples on lists, dictionaries,
strings, etc. . .
In [38]: variable_args('one', 'two', x=1, y=2, z=3)
args is ('one', 'two') 2.4.10 Exercises
kwargs is {'x': 1, 'y': 2, 'z': 3}

Exercise: Fibonacci sequence

2.4.7 Docstrings
Write a function that displays the n first terms of the Fibonacci sequence, defined by:
Documentation about what the function does and its parameters. General convention: ⎧
⎨ 𝑈0 = 0
In [39]: def funcname(params): 𝑈1 = 1
....: """Concise one-line sentence describing the function. ⎩
𝑈𝑛+2 = 𝑈𝑛+1 + 𝑈𝑛
....:
....: Extended summary which can contain multiple paragraphs.
....: """
....: # function body Exercise: Quicksort
....: pass
Implement the quicksort algorithm, as defined by wikipedia
....:

(continues on next page)

2.4. Defining functions 30 2.4. Defining functions 31

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

function quicksort(array)  Tip

var list less, greater
if length(array) < 2 Other interpreters also offer the possibility to execute scripts (e.g., execfile in the plain Python
return array interpreter, etc.).
select and remove a pivot value pivot from array
for each x in array
if x < pivot + 1 then append x to less It is also possible In order to execute this script as a standalone program, by executing the script inside
else append x to greater a shell terminal (Linux/Mac console or cmd Windows console). For example, if we are in the same
return concatenate(quicksort(less), pivot, quicksort(greater)) directory as the test.py file, we can execute this in a console:
$ python test.py
Hello
how
2.5 Reusing code: scripts and modules are
you?
For now, we have typed all instructions in the interpreter. For longer sets of instructions we need to
change track and write the code in text files (using a text editor), that we will call either scripts or
modules. Use your favorite text editor (provided it offers syntax highlighting for Python), or the editor  Tip
that comes with the Scientific Python Suite you may be using.
Standalone scripts may also take command-line arguments

2.5.1 Scripts In file.py:

import sys
print(sys.argv)
 Tip

Let us first write a script, that is a file with a sequence of instructions that are executed each time $ python file.py test arguments
the script is called. Instructions may be e.g. copied-and-pasted from the interpreter (but take care ['file.py', 'test', 'arguments']
to respect indentation rules!).
. Warning
The extension for Python files is .py. Write or copy-and-paste the following lines in a file called test.py
Don’t implement option parsing yourself. Use a dedicated module such as argparse.
message = "Hello how are you?"
for word in message.split():
print(word)
2.5.2 Importing objects from modules
In [3]: import os
 Tip
In [4]: os
Let us now execute the script interactively, that is inside the Ipython interpreter. This is maybe the Out[4]: <module 'os' (frozen)>
most common use of scripts in scientific computing.
In [5]: os.listdir('.')
Out[5]:
['profile_1f1y_tei',
ò Note
'profile_fafmr45r',
in Ipython, the syntax to execute a script is %run script.py. For example, 'profile__degnych',
'clr-debug-pipe-1744-11435-out',
'dotnet-diagnostic-1744-11435-socket',
In [1]: %run test.py 'systemd-private-4d98e5f0d5f9404e9f14cc2cacc4c564-chrony.service-EWImxL',
Hello 'profile_z1bv__o8',
how 'profile_piq4dc_g',
are 'profile_vpgc00e4',
you? 'profile_clzffj76',
'clr-debug-pipe-841-947-in',
In [2]: message 'profile_pk7qvgfa',
Out[2]: 'Hello how are you?' 'profile_r5e9rl10',
'profile_xzyebo3o',
The script has been executed. Moreover the variables defined in the script (such as message) are now 'snap-private-tmp',
available inside the interpreter’s namespace. (continues on next page)

2.5. Reusing code: scripts and modules 32 2.5. Reusing code: scripts and modules 33
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

'profile_mgqqezet', 'profile_73odl5dt',
'profile_orscovps', 'profile_btzppd3p',
'profile_9k60vb8a', 'profile_9j0h7xls',
'profile_9fkghju_', 'profile_ozgkewko',
'profile_zviham2g', 'profile_59z5oe0h',
'profile__s62lnm2', 'profile_sczqnqvw',
'.font-unix', 'profile_le2se4kz',
'profile_crltuoda', 'profile_03ywiqkf',
'profile_dozs3xo_', 'profile_y0boqv14']
'dotnet-diagnostic-841-947-socket',
'profile_23ip7ee8', And also:
'profile_i3f29kbe',
'systemd-private-4d98e5f0d5f9404e9f14cc2cacc4c564-ModemManager.service-qufeDJ', In [6]: from os import listdir
'systemd-private-4d98e5f0d5f9404e9f14cc2cacc4c564-polkit.service-xGYWnt',
'clr-debug-pipe-841-947-out', Importing shorthands:
'profile_dyxf7nzz',
'.XIM-unix', In [7]: import numpy as np
'.ICE-unix',
'profile_hqjnr6r0',
'profile_tikc0z65', . Warning
'profile_syb_pmuh',
'profile_j65t20ng', from os import *
'profile_sxsm9anp',
This is called the star import and please, Do not use it
'profile_0rj8q0f2',
'profile_e_og5pja', • Makes the code harder to read and understand: where do symbols come from?
'profile_d6zq9bwn',
• Makes it impossible to guess the functionality by the context and the name (hint: os.name is
'dotnet-diagnostic-1727-11313-socket',
the name of the OS), and to profit usefully from tab completion.
'systemd-private-4d98e5f0d5f9404e9f14cc2cacc4c564-haveged.service-FVI9vk',
'profile_bcpbnmci', • Restricts the variable names you can use: os.name might override name, or vise-versa.
'profile_bb5we1ns',
• Creates possible name clashes between modules.
'profile_frf9qh57',
'profile_225m_ihr', • Makes the code impossible to statically check for undefined symbols.
'profile_u0lm7yr8',
'profile_u784oe5n',
'profile_ujh340a4',
'.X11-unix',  Tip
'hsperfdata_root',
'profile_4b4m_zxn', Modules are thus a good way to organize code in a hierarchical way. Actually, all the scientific
'clr-debug-pipe-1727-11313-in', computing tools we are going to use are modules:
'profile_sxshlppf', >>> import numpy as np # data arrays
'profile__1jotx1e', >>> np.linspace(0, 10, 6)
'systemd-private-4d98e5f0d5f9404e9f14cc2cacc4c564-systemd-logind.service-hWNIpv', array([ 0., 2., 4., 6., 8., 10.])
'clr-debug-pipe-1744-11435-in', >>> import scipy as sp # scientific computing
'profile_eib7rssh',
'profile_gbzts59j',
'profile_nh1ihwo0',
'profile_y1sgvu4l',
2.5.3 Creating modules
'profile_hcucx43r',
'profile_k1n6qi_w',  Tip
'clr-debug-pipe-1727-11313-out',
'profile_2ag5tpgo', If we want to write larger and better organized programs (compared to simple scripts), where some
'profile__ot4ykxk', objects are defined, (variables, functions, classes) and that we want to reuse several times, we have
'profile_3_u_72gw', to create our own modules.
'profile_eoap_hci',
'profile_kp5b28ji',
Let us create a module demo contained in the file demo.py:
'systemd-private-4d98e5f0d5f9404e9f14cc2cacc4c564-systemd-resolved.service-HcmpIM',
(continues on next page)

2.5. Reusing code: scripts and modules 34 2.5. Reusing code: scripts and modules 35
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

"A demo module."
['__builtins__',
'__doc__',
def print_b(): '__file__',
"Prints b." '__name__',
print("b") '__package__',
'c',
'd',
def print_a(): 'print_a',
"Prints a." 'print_b']
print("a")
In [15]: demo.<TAB>
demo.c demo.print_a demo.py
c = 2 demo.d demo.print_b demo.pyc
d = 2
Importing objects from modules into the main namespace
In [16]: from demo import print_a, print_b
 Tip
In [17]: whos
In this file, we defined two functions print_a and print_b. Suppose we want to call the print_a
Variable Type Data/Info
function from the interpreter. We could execute the file as a script, but since we just want to have
--------------------------------
access to the function print_a, we are rather going to import it as a module. The syntax is as
demo module <module 'demo' from 'demo.py'>
follows.
print_a function <function print_a at 0xb7421534>
print_b function <function print_b at 0xb74214c4>
In [8]: import demo
In [18]: print_a()
In [9]: demo.print_a() a
a

In [10]: demo.print_b() . Warning

b
Module caching
Importing the module gives access to its objects, using the module.object syntax. Don’t forget to put Modules are cached: if you modify demo.py and re-import it in the old session, you will
the module’s name before the object’s name, otherwise Python won’t recognize the instruction. get the old one.
Introspection Solution:
In [11]: demo? In [10]: importlib.reload(demo)
Type: module
Base Class: <type 'module'>
String Form: <module 'demo' from 'demo.py'> 2.5.4 ‘__main__’ and module loading
Namespace: Interactive
File: /home/varoquau/Projects/Python_talks/scipy_2009_tutorial/source/
˓→demo.py  Tip
Docstring:
A demo module. Sometimes we want code to be executed when a module is run directly, but not when it is imported
by another module. if __name__ == '__main__' allows us to check whether the module is being
In [12]: who run directly.
demo
File demo2.py:
In [13]: whos
Variable Type Data/Info def print_b():
------------------------------ "Prints b."
demo module <module 'demo' from 'demo.py'> print("b")

In [14]: dir(demo)
Out[14]: def print_a():
(continues on next page) (continues on next page)

2.5. Reusing code: scripts and modules 36 2.5. Reusing code: scripts and modules 37
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) • write your own modules within directories already defined in the search path (e.g. $HOME/.venv/
"Prints a." lectures/lib64/python3.11/site-packages). You may use symbolic links (on Linux) to keep
print("a") the code somewhere else.
• modify the environment variable PYTHONPATH to include the directories containing the user-defined
modules.
# print_b() runs on import
print_b()
 Tip
if __name__ == "__main__":
# print_a() is only executed when the module is run directly. On Linux/Unix, add the following line to a file read by the shell at startup (e.g. /etc/profile,
print_a() .profile)
export PYTHONPATH=$PYTHONPATH:/home/emma/user_defined_modules
Importing it:
On Windows, https://support.microsoft.com/kb/310519 explains how to handle environment
In [19]: import demo2 variables.
b

In [20]: import demo2 • or modify the sys.path variable itself within a Python script.

Running it:  Tip

In [21]: %run demo2
import sys
b
new_path = '/home/emma/user_defined_modules'
a
if new_path not in sys.path:
sys.path.append(new_path)
2.5.5 Scripts or modules? How to organize your code This method is not very robust, however, because it makes the code less portable (user-
dependent path) and because you have to add the directory to your sys.path each time you
ò Note want to import from a module in this directory.

Rule of thumb
• Sets of instructions that are called several times should be written inside functions for better ã See also
code reusability.
See https://docs.python.org/3/tutorial/modules.html for more information about modules.
• Functions (or other bits of code) that are called from several scripts should be written inside
a module, so that only the module is imported in the different scripts (do not copy-and-paste
your functions in the different scripts!).
2.5.6 Packages
A directory that contains many modules is called a package. A package is a module with submodules
How modules are found and imported (which can have submodules themselves, etc.). A special file called __init__.py (which may be empty)
tells Python that the directory is a Python package, from which modules can be imported.
When the import mymodule statement is executed, the module mymodule is searched in a given list of
directories. This list includes a list of installation-dependent default path (e.g., /usr/lib64/python3. $ ls
11) as well as the list of directories specified by the environment variable PYTHONPATH. _build_utils/ fft/ _lib/ odr/ spatial/
The list of directories searched by Python is given by the sys.path variable cluster/ fftpack/ linalg/ optimize/ special/
conftest.py __init__.py linalg.pxd optimize.pxd special.pxd
In [22]: import sys constants/ integrate/ meson.build setup.py stats/
datasets/ interpolate/ misc/ signal/
In [23]: sys.path _distributor_init.py io/ ndimage/ sparse/
Out[23]: $ cd ndimage
['/home/runner/work/scientific-python-lectures/scientific-python-lectures', $ ls
'/opt/hostedtoolcache/Python/3.12.9/x64/lib/python312.zip', _filters.py __init__.py _measurements.py morphology.py src/
'/opt/hostedtoolcache/Python/3.12.9/x64/lib/python3.12', filters.py _interpolation.py measurements.py _ni_docstrings.py tests/
'/opt/hostedtoolcache/Python/3.12.9/x64/lib/python3.12/lib-dynload', _fourier.py interpolation.py meson.build _ni_support.py utils/
'/opt/hostedtoolcache/Python/3.12.9/x64/lib/python3.12/site-packages'] fourier.py LICENSE.txt _morphology.py setup.py

Modules must be located in the search path, therefore you can: From Ipython:

2.5. Reusing code: scripts and modules 38 2.5. Reusing code: scripts and modules 39
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Spaces
In [24]: import scipy as sp
Write well-spaced code: put whitespaces after commas, around arithmetic operators, etc.:
In [25]: sp.__file__
Out[25]: '/opt/hostedtoolcache/Python/3.12.9/x64/lib/python3.12/site-packages/scipy/__ >>> a = 1 # yes
˓→init__.py'
>>> a=1 # too cramped

In [26]: sp.version.version A certain number of rules for writing “beautiful” code (and more importantly using the same
Out[26]: '1.15.2' conventions as anybody else!) are given in the Style Guide for Python Code.

In [27]: sp.ndimage.morphology.binary_dilation?
Signature:
sp.ndimage.morphology.binary_dilation( Quick read
input,
If you want to do a first quick pass through the Scientific Python Lectures to learn the ecosystem,
structure=None,
you can directly skip to the next chapter: NumPy: creating and manipulating numerical data.
iterations=1,
mask=None, The remainder of this chapter is not necessary to follow the rest of the intro part. But be sure to
output=None, come back and finish this chapter later.
border_value=0,
origin=0,
brute_force=False,
)
Docstring: 2.6 Input and Output
Multidimensional binary dilation with the given structuring element.
... To be exhaustive, here are some information about input and output in Python. Since we will use the
NumPy methods to read and write files, you may skip this chapter at first reading.
We write or read strings to/from files (other types must be converted to strings). To write in a file:
2.5.7 Good practices
• Use meaningful object names >>> f = open('workfile', 'w') # opens the workfile file
>>> type(f)
• Indentation: no choice! <class '_io.TextIOWrapper'>
>>> f.write('This is a test \nand another test')
>>> f.close()
 Tip

Indenting is compulsory in Python! Every command block following a colon bears an additional To read from a file
indentation level with respect to the previous line with a colon. One must therefore indent after
In [1]: f = open('workfile', 'r')
def f(): or while:. At the end of such logical blocks, one decreases the indentation depth
(and re-increases it if a new block is entered, etc.)
In [2]: s = f.read()
Strict respect of indentation is the price to pay for getting rid of { or ; characters that delineate
logical blocks in other languages. Improper indentation leads to errors such as In [3]: print(s)
This is a test
------------------------------------------------------------
and another test
IndentationError: unexpected indent (test.py, line 2)

All this indentation business can be a bit confusing in the beginning. However, with the clear In [4]: f.close()
indentation, and in the absence of extra characters, the resulting code is very nice to read
compared to other languages.
ã See also
• Indentation depth: Inside your text editor, you may choose to indent with any positive number For more details: https://docs.python.org/3/tutorial/inputoutput.html
of spaces (1, 2, 3, 4, . . . ). However, it is considered good practice to indent with 4 spaces. You
may configure your editor to map the Tab key to a 4-space indentation.
• Style guidelines 2.6.1 Iterating over a file
Long lines: you should not write very long lines that span over more than (e.g.) 80 characters. In [5]: f = open('workfile', 'r')
Long lines can be broken with the \ character
In [6]: for line in f:
>>> long_line = "Here is a very very long line \ ...: print(line)
... that we break in two parts." (continues on next page)

2.5. Reusing code: scripts and modules 40 2.6. Input and Output 41
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

...: 'profile_piq4dc_g',
This is a test 'profile_vpgc00e4',
and another test 'profile_z5azo0zp',
'profile_clzffj76',
In [7]: f.close() 'clr-debug-pipe-841-947-in',
'profile_zh2mx3mh',
'profile_pk7qvgfa',
File modes 'profile_r5e9rl10',
• Read-only: r 'profile_xzyebo3o',
'snap-private-tmp',
• Write-only: w 'profile_mgqqezet',
– Note: Create a new file or overwrite existing file. 'profile_orscovps',
'profile_9k60vb8a',
• Append a file: a 'profile_9fkghju_',
• Read and Write: r+ 'profile_zviham2g',
'profile__s62lnm2',
• Binary mode: b '.font-unix',
– Note: Use for binary files, especially on Windows. 'profile_crltuoda',
'profile_dozs3xo_',
'dotnet-diagnostic-841-947-socket',
'profile_23ip7ee8',
'profile_i3f29kbe',
2.7 Standard Library 'profile_y6wv83ue',
'systemd-private-4d98e5f0d5f9404e9f14cc2cacc4c564-ModemManager.service-qufeDJ',
'systemd-private-4d98e5f0d5f9404e9f14cc2cacc4c564-polkit.service-xGYWnt',
ò Note
'clr-debug-pipe-841-947-out',
Reference document for this section: 'profile_dyxf7nzz',
'.XIM-unix',
• The Python Standard Library documentation: https://docs.python.org/3/library/index.html '.ICE-unix',
• Python Essential Reference, David Beazley, Addison-Wesley Professional 'profile_hqjnr6r0',
'profile_tikc0z65',
'profile_de74ztqi',
2.7.1 os module: operating system functionality 'profile_syb_pmuh',
'profile_j65t20ng',
“A portable way of using operating system dependent functionality.” 'profile_sxsm9anp',
'profile_0rj8q0f2',
Directory and file manipulation 'profile_e_og5pja',
'profile_o1l4z05e',
Current directory:
'profile_d6zq9bwn',
In [1]: import os 'dotnet-diagnostic-1727-11313-socket',
'systemd-private-4d98e5f0d5f9404e9f14cc2cacc4c564-haveged.service-FVI9vk',
In [2]: os.getcwd() 'profile_bcpbnmci',
Out[2]: '/tmp' 'profile_bb5we1ns',
'profile_frf9qh57',
List a directory: 'profile_225m_ihr',
'profile_u0lm7yr8',
In [3]: os.listdir(os.curdir) 'profile_u784oe5n',
Out[3]: 'profile_ujh340a4',
['profile_1f1y_tei', '.X11-unix',
'profile_fafmr45r', 'hsperfdata_root',
'profile__degnych', 'profile_4b4m_zxn',
'clr-debug-pipe-1744-11435-out', 'profile_u7_4sy2e',
'dotnet-diagnostic-1744-11435-socket', 'clr-debug-pipe-1727-11313-in',
'systemd-private-4d98e5f0d5f9404e9f14cc2cacc4c564-chrony.service-EWImxL', 'profile_sxshlppf',
'profile_z1bv__o8', 'profile__1jotx1e',
'profile_eyv5c5ub', 'profile_bg63e2j6',
(continues on next page) (continues on next page)

2.7. Standard Library 42 2.7. Standard Library 43

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

'systemd-private-4d98e5f0d5f9404e9f14cc2cacc4c564-systemd-logind.service-hWNIpv', Out[13]: True
'clr-debug-pipe-1744-11435-in',
'profile_eib7rssh', In [14]: os.remove('junk.txt')
'profile_gbzts59j',
'profile_nh1ihwo0', In [15]: 'junk.txt' in os.listdir(os.curdir)
'profile_y1sgvu4l', Out[15]: False
'profile_hcucx43r',
'profile_k1n6qi_w',
'profile_bk1ic5bw', os.path: path manipulations
'clr-debug-pipe-1727-11313-out', os.path provides common operations on pathnames.
'profile_2ag5tpgo',
'profile__ot4ykxk', In [16]: fp = open('junk.txt', 'w')
'profile_3_u_72gw',
'profile_ufq68h_c', In [17]: fp.close()
'profile_eoap_hci',
'profile_kp5b28ji', In [18]: a = os.path.abspath('junk.txt')
'systemd-private-4d98e5f0d5f9404e9f14cc2cacc4c564-systemd-resolved.service-HcmpIM',
'profile_73odl5dt', In [19]: a
'profile_btzppd3p', Out[19]: '/tmp/junk.txt'
'profile_9j0h7xls',
'profile_ozgkewko', In [20]: os.path.split(a)
'profile_59z5oe0h', Out[20]: ('/tmp', 'junk.txt')
'profile_sczqnqvw',
'profile_le2se4kz', In [21]: os.path.dirname(a)
'profile_03ywiqkf', Out[21]: '/tmp'
'profile_y0boqv14',
'profile_uk3dqt55'] In [22]: os.path.basename(a)
Out[22]: 'junk.txt'
Make a directory:
In [23]: os.path.splitext(os.path.basename(a))
In [4]: os.mkdir('junkdir') Out[23]: ('junk', '.txt')

In [5]: 'junkdir' in os.listdir(os.curdir) In [24]: os.path.exists('junk.txt')

Out[5]: True Out[24]: True

Rename the directory: In [25]: os.path.isfile('junk.txt')

Out[25]: True
In [6]: os.rename('junkdir', 'foodir')
In [26]: os.path.isdir('junk.txt')
In [7]: 'junkdir' in os.listdir(os.curdir)
Out[26]: False
Out[7]: False
In [27]: os.path.expanduser('~/local')
In [8]: 'foodir' in os.listdir(os.curdir)
Out[27]: '/home/runner/local'
Out[8]: True
In [28]: os.path.join(os.path.expanduser('~'), 'local', 'bin')
In [9]: os.rmdir('foodir')
Out[28]: '/home/runner/local/bin'
In [10]: 'foodir' in os.listdir(os.curdir)
Out[10]: False Running an external command

Delete a file: In [29]: os.system('ls')

Out[29]: 0
In [11]: fp = open('junk.txt', 'w')

In [12]: fp.close()
ò Note
In [13]: 'junk.txt' in os.listdir(os.curdir) Alternative to os.system
(continues on next page)

2.7. Standard Library 44 2.7. Standard Library 45

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

A noteworthy alternative to os.system is the sh module. Which provides much more convenient
ways to obtain the output, error stream and exit code of the external command. /tmp/README
/tmp/README
In [30]: import sh
/tmp/README
In [31]: com = sh.ls()
/tmp/README
/tmp/README
In [31]: print(com)
/tmp/README
basic_types.rst exceptions.rst oop.rst standard_library.rst
/tmp/README
control_flow.rst first_steps.rst python_language.rst
/tmp/README
demo2.py functions.rst python-logo.png
/tmp/README
demo.py io.rst reusing_code.rst
/tmp/README
/tmp/README
In [32]: type(com)
/tmp/README
Out[32]: str
/tmp/README
/tmp/README
/tmp/README
Walking a directory
/tmp/README
os.path.walk generates a list of filenames in a directory tree. /tmp/README
/tmp/README
In [33]: for dirpath, dirnames, filenames in os.walk(os.curdir): /tmp/README
....: for fp in filenames: /tmp/README
....: print(os.path.abspath(fp)) /tmp/README
....: /tmp/README
....: /tmp/README
/tmp/clr-debug-pipe-1744-11435-out /tmp/README
/tmp/dotnet-diagnostic-1744-11435-socket /tmp/README
/tmp/clr-debug-pipe-841-947-in /tmp/README
/tmp/dotnet-diagnostic-841-947-socket /tmp/README
/tmp/junk.txt /tmp/README
/tmp/clr-debug-pipe-841-947-out /tmp/README
/tmp/dotnet-diagnostic-1727-11313-socket /tmp/README
/tmp/clr-debug-pipe-1727-11313-in /tmp/README
/tmp/clr-debug-pipe-1744-11435-in /tmp/README
/tmp/clr-debug-pipe-1727-11313-out /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/bookmarks /tmp/README
/tmp/dhist /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
/tmp/README /tmp/README
(continues on next page)

2.7. Standard Library 46 2.7. Standard Library 47

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Environment variables:
In [43]: sys.path
In [34]: os.environ.keys() Out[43]:
Out[34]: KeysView(environ({'SHELL': '/bin/bash', 'COLORTERM': 'truecolor', ...})) ['/home/runner/work/scientific-python-lectures/scientific-python-lectures',
'/opt/hostedtoolcache/Python/3.12.9/x64/lib/python312.zip',
In [35]: os.environ['SHELL'] '/opt/hostedtoolcache/Python/3.12.9/x64/lib/python3.12',
Out[35]: '/bin/bash' '/opt/hostedtoolcache/Python/3.12.9/x64/lib/python3.12/lib-dynload',
'/opt/hostedtoolcache/Python/3.12.9/x64/lib/python3.12/site-packages']

2.7.2 shutil: high-level file operations

2.7.5 pickle: easy persistence
The shutil provides useful file operations:
Useful to store arbitrary objects to a file. Not safe or fast!
• shutil.rmtree: Recursively delete a directory tree.
In [44]: import pickle
• shutil.move: Recursively move a file or directory to another location.
• shutil.copy: Copy files or directories. In [45]: l = [1, None, 'Stan']

2.7.3 glob: Pattern matching on files In [46]: with open('test.pkl', 'wb') as file:
....: pickle.dump(l, file)
The glob module provides convenient file pattern matching. ....:
Find all files ending in .txt:
In [47]: with open('test.pkl', 'rb') as file:
In [36]: import glob ....: out = pickle.load(file)
....:
In [37]: glob.glob('*.txt')
Out[37]: ['junk.txt'] In [48]: out
Out[48]: [1, None, 'Stan']

2.7.4 sys module: system-specific information

System-specific information related to the Python interpreter. Exercise

• Which version of python are you running and where is it installed: Write a program to search your PYTHONPATH for the module site.py.

In [38]: import sys path_site

In [39]: sys.platform
Out[39]: 'linux'

In [40]: sys.version 2.8 Exception handling in Python

Out[40]: '3.12.9 (main, Feb 5 2025, 03:22:31) [GCC 13.3.0]'
It is likely that you have raised Exceptions if you have typed all the previous commands of the tutorial.
For example, you may have raised an exception if you entered a command with a typo.
In [41]: sys.prefix
Out[41]: '/opt/hostedtoolcache/Python/3.12.9/x64' Exceptions are raised by different kinds of errors arising when executing Python code. In your own code,
you may also catch errors, or define custom error types. You may want to look at the descriptions of the
• List of command line arguments passed to a Python script: the built-in Exceptions when looking for the right exception type.

In [42]: sys.argv
Out[42]:
2.8.1 Exceptions
['/opt/hostedtoolcache/Python/3.12.9/x64/lib/python3.12/site-packages/sphinx/__main__. Exceptions are raised by errors in Python:
˓→py',

'-b', In [1]: 1/0

'latex', ---------------------------------------------------------------------------
'-d', ZeroDivisionError Traceback (most recent call last)
'build/doctrees', Cell In[1], line 1
'.', ----> 1 1/0
'build/latex']
ZeroDivisionError: division by zero
sys.path is a list of strings that specifies the search path for modules. Initialized from PYTHONPATH:
(continues on next page)

2.7. Standard Library 48 2.8. Exception handling in Python 49

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) try/finally

In [2]: 1 + 'e'
In [10]: try:
---------------------------------------------------------------------------
....: x = int(input('Please enter a number: '))
TypeError Traceback (most recent call last)
....: finally:
Cell In[2], line 1
....: print('Thank you for your input')
----> 1 1 + 'e'
....:
Please enter a number: a
TypeError: unsupported operand type(s) for +: 'int' and 'str'
Thank you for your input
---------------------------------------------------------------------------
In [3]: d = {1:1, 2:2}
ValueError Traceback (most recent call last)
Cell In[10], line 2
In [4]: d[3]
1 try:
---------------------------------------------------------------------------
----> 2 x = int(input('Please enter a number: '))
KeyError Traceback (most recent call last)
3 finally:
Cell In[4], line 1
4 print('Thank you for your input')
----> 1 d[3]
ValueError: invalid literal for int() with base 10: 'a'
KeyError: 3
Important for resource management (e.g. closing a file)
In [5]: l = [1, 2, 3]
Easier to ask for forgiveness than for permission
In [6]: l[4]
--------------------------------------------------------------------------- In [11]: def print_sorted(collection):
IndexError Traceback (most recent call last) ....: try:
Cell In[6], line 1 ....: collection.sort()
----> 1 l[4] ....: except AttributeError:
....: pass # The pass statement does nothing
IndexError: list index out of range ....: print(collection)
....:
In [7]: l.foobar
--------------------------------------------------------------------------- In [12]: print_sorted([1, 3, 2])
AttributeError Traceback (most recent call last) [1, 2, 3]
Cell In[7], line 1
----> 1 l.foobar In [13]: print_sorted(set((1, 3, 2)))
{1, 2, 3}
AttributeError: 'list' object has no attribute 'foobar'
In [14]: print_sorted('132')
132
As you can see, there are different types of exceptions for different errors.

2.8.2 Catching exceptions 2.8.3 Raising exceptions

try/except • Capturing and reraising an exception:

In [8]: while True: In [15]: def filter_name(name):

...: try: ....: try:
...: x = int(input('Please enter a number: ')) ....: name = name.encode('ascii')
...: break ....: except UnicodeError as e:
...: except ValueError: ....: if name == 'Gaël':
...: print('That was no valid number. Try again...') ....: print('OK, Gaël')
...: ....: else:
Please enter a number: a ....: raise e
That was no valid number. Try again... ....: return name
Please enter a number: 1 ....:

In [9]: x In [16]: filter_name('Gaël')

Out[9]: 1 OK, Gaël
Out[16]: 'Gaël'
(continues on next page)

2.8. Exception handling in Python 50 2.8. Exception handling in Python 51

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (methods) and variables (attributes), we will be able to use:

>>> class Student(object):

In [17]: filter_name('Stéfan')
... def __init__(self, name):
---------------------------------------------------------------------------
... self.name = name
UnicodeEncodeError Traceback (most recent call last)
... def set_age(self, age):
Cell In[17], line 1
... self.age = age
----> 1 filter_name('Stéfan')
... def set_major(self, major):
... self.major = major
Cell In[15], line 8, in filter_name(name)
...
6 print('OK, Gaël')
>>> anna = Student('anna')
7 else:
>>> anna.set_age(21)
----> 8 raise e
>>> anna.set_major('physics')
9 return name

Cell In[15], line 3, in filter_name(name) In the previous example, the Student class has __init__, set_age and set_major methods. Its at-
1 def filter_name(name): tributes are name, age and major. We can call these methods and attributes with the following notation:
2 try: classinstance.method or classinstance.attribute. The __init__ constructor is a special method
----> 3 name = name.encode('ascii') we call with: MyClass(init parameters if any).
4 except UnicodeError as e: Now, suppose we want to create a new class MasterStudent with the same methods and attributes as
5 if name == 'Gaël': the previous one, but with an additional internship attribute. We won’t copy the previous class, but
inherit from it:
UnicodeEncodeError: 'ascii' codec can't encode character '\xe9' in position 2:␣
˓→ordinal not in range(128) >>> class MasterStudent(Student):
... internship = 'mandatory, from March to June'
• Exceptions to pass messages between parts of the code: ...
>>> james = MasterStudent('james')
In [18]: def achilles_arrow(x): >>> james.internship
....: if abs(x - 1) < 1e-3: 'mandatory, from March to June'
....: raise StopIteration >>> james.set_age(23)
....: x = 1 - (1-x)/2. >>> james.age
....: return x 23
....:
The MasterStudent class inherited from the Student attributes and methods.
In [19]: x = 0
Thanks to classes and object-oriented programming, we can organize code with different classes corre-
In [20]: while True: sponding to different objects we encounter (an Experiment class, an Image class, a Flow class, etc.), with
....: try: their own methods and attributes. Then we can use inheritance to consider variations around a base
....: x = achilles_arrow(x) class and reuse code. Ex : from a Flow base class, we can create derived StokesFlow, TurbulentFlow,
....: except StopIteration: PotentialFlow, etc.
....: break
....:
....:

In [21]: x
Out[21]: 0.9990234375

Use exceptions to notify certain conditions are met (e.g. StopIteration) or not (e.g. custom error raising)

2.9 Object-oriented programming (OOP)

Python supports object-oriented programming (OOP). The goals of OOP are:
• to organize the code, and
• to reuse code in similar contexts.
Here is a small example: we create a Student class, which is an object gathering several custom functions

2.9. Object-oriented programming (OOP) 52 2.9. Object-oriented programming (OOP) 53

 Scientific Python Lectures, Edition 2025.1rc0.dev0

3.1.1 What are NumPy and NumPy arrays?

NumPy arrays
Python objects
• high-level number objects: integers, floating point

3
• containers: lists (costless insertion and append), dictionaries (fast lookup)
NumPy provides
• extension package to Python for multi-dimensional arrays
• closer to hardware (efficiency)
CHAPTER
• designed for scientific computation (convenience)
• Also known as array oriented computing

NumPy: creating and manipulating

>>> import numpy as np
>>> a = np.array([0, 1, 2, 3])
>>> a
numerical data array([0, 1, 2, 3])

 Tip

For example, An array containing:

• values of an experiment/simulation at discrete time steps
• signal recorded by a measurement device, e.g. sound wave
Authors: Emmanuelle Gouillart, Didrik Pinte, Gaël Varoquaux, and Pauli Virtanen
• pixels of an image, grey-level or colour
This chapter gives an overview of NumPy, the core tool for performant numerical computing with Python.
• 3-D data measured at different X-Y-Z positions, e.g. MRI scan
• ...

3.1 The NumPy array object Why it is useful: Memory-efficient container that provides fast numerical operations.

In [1]: L = range(1000)
Section contents
In [2]: %timeit [i**2 for i in L]
• What are NumPy and NumPy arrays? 53.9 us +- 961 ns per loop (mean +- std. dev. of 7 runs, 10,000 loops each)
• Creating arrays
In [3]: a = np.arange(1000)
• Basic data types
• Basic visualization In [4]: %timeit a**2
950 ns +- 4.13 ns per loop (mean +- std. dev. of 7 runs, 1,000,000 loops each)
• Indexing and slicing
• Copies and views NumPy Reference documentation
• Fancy indexing
• On the web: https://numpy.org/doc/
• Interactive help:

In [5]: np.array?
Docstring:
(continues on next page)

54 3.1. The NumPy array object 55

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

array(object, dtype=None, *, copy=True, order='K', subok=False, ndmin=0,
like=None) .. versionadded:: 1.20.0

Create an array. Returns

-------
Parameters out : ndarray
---------- An array object satisfying the specified requirements.
object : array_like
An array, any object exposing the array interface, an object whose See Also
``__array__`` method returns an array, or any (nested) sequence. --------
If object is a scalar, a 0-dimensional array containing object is empty_like : Return an empty array with shape and type of input.
returned. ones_like : Return an array of ones with shape and type of input.
dtype : data-type, optional zeros_like : Return an array of zeros with shape and type of input.
The desired data-type for the array. If not given, NumPy will try to use full_like : Return a new array with shape of input filled with value.
a default ``dtype`` that can represent the values (by applying promotion empty : Return a new uninitialized array.
rules when necessary.) ones : Return a new array setting values to one.
copy : bool, optional zeros : Return a new array setting values to zero.
If ``True`` (default), then the array data is copied. If ``None``, full : Return a new array of given shape filled with value.
a copy will only be made if ``__array__`` returns a copy, if obj is copy: Return an array copy of the given object.
a nested sequence, or if a copy is needed to satisfy any of the other
requirements (``dtype``, ``order``, etc.). Note that any copy of
the data is shallow, i.e., for arrays with object dtype, the new Notes
array will point to the same objects. See Examples for `ndarray.copy`. -----
For ``False`` it raises a ``ValueError`` if a copy cannot be avoided. When order is 'A' and ``object`` is an array in neither 'C' nor 'F' order,
Default: ``True``. and a copy is forced by a change in dtype, then the order of the result is
order : {'K', 'A', 'C', 'F'}, optional not necessarily 'C' as expected. This is likely a bug.
Specify the memory layout of the array. If object is not an array, the
newly created array will be in C order (row major) unless 'F' is Examples
specified, in which case it will be in Fortran order (column major). --------
If object is an array the following holds. >>> import numpy as np
>>> np.array([1, 2, 3])
===== ========= =================================================== array([1, 2, 3])
order no copy copy=True
===== ========= =================================================== Upcasting:
'K' unchanged F & C order preserved, otherwise most similar order
'A' unchanged F order if input is F and not C, otherwise C order >>> np.array([1, 2, 3.0])
'C' C order C order array([ 1., 2., 3.])
'F' F order F order
===== ========= =================================================== More than one dimension:

When ``copy=None`` and a copy is made for other reasons, the result is >>> np.array([[1, 2], [3, 4]])
the same as if ``copy=True``, with some exceptions for 'A', see the array([[1, 2],
Notes section. The default order is 'K'. [3, 4]])
subok : bool, optional
If True, then sub-classes will be passed-through, otherwise Minimum dimensions 2:
the returned array will be forced to be a base-class array (default).
ndmin : int, optional >>> np.array([1, 2, 3], ndmin=2)
Specifies the minimum number of dimensions that the resulting array([[1, 2, 3]])
array should have. Ones will be prepended to the shape as
needed to meet this requirement. Type provided:
like : array_like, optional
Reference object to allow the creation of arrays which are not >>> np.array([1, 2, 3], dtype=complex)
NumPy arrays. If an array-like passed in as ``like`` supports array([ 1.+0.j, 2.+0.j, 3.+0.j])
the ``__array_function__`` protocol, the result will be defined
by it. In this case, it ensures the creation of an array object Data-type consisting of more than one element:
compatible with that passed in via this argument.
(continues on next page) (continues on next page)

3.1. The NumPy array object 56 3.1. The NumPy array object 57
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> b = np.array([[0, 1, 2], [3, 4, 5]]) # 2 x 3 array
>>> x = np.array([(1,2),(3,4)],dtype=[('a','<i4'),('b','<i4')]) >>> b
>>> x['a'] array([[0, 1, 2],
array([1, 3], dtype=int32) [3, 4, 5]])
>>> b.ndim
Creating an array from sub-classes: 2
>>> b.shape
>>> np.array(np.asmatrix('1 2; 3 4')) (2, 3)
array([[1, 2], >>> len(b) # returns the size of the first dimension
[3, 4]]) 2

>>> np.array(np.asmatrix('1 2; 3 4'), subok=True) >>> c = np.array([[[1], [2]], [[3], [4]]])

matrix([[1, 2], >>> c
[3, 4]]) array([[[1],
Type: builtin_function_or_method [2]],

[[3],
 Tip [4]]])
>>> c.shape
>>> help(np.array) (2, 2, 1)
Help on built-in function array in module numpy:

array(...) Exercise: Simple arrays

array(object, dtype=None, ...
• Create a simple two dimensional array. First, redo the examples from above. And then create
your own: how about odd numbers counting backwards on the first row, and even numbers on
• Looking for something: the second?
In [6]: np.con*? • Use the functions len(), numpy.shape() on these arrays. How do they relate to each other?
np.concat And to the ndim attribute of the arrays?
np.concatenate
np.conj
np.conjugate Functions for creating arrays
np.convolve
 Tip
Import conventions In practice, we rarely enter items one by one. . .
The recommended convention to import NumPy is:

>>> import numpy as np • Evenly spaced:

>>> a = np.arange(10) # 0 .. n-1 (!)

>>> a
3.1.2 Creating arrays
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
Manual construction of arrays >>> b = np.arange(1, 9, 2) # start, end (exclusive), step
>>> b
• 1-D: array([1, 3, 5, 7])
>>> a = np.array([0, 1, 2, 3])
>>> a • or by number of points:
array([0, 1, 2, 3]) >>> c = np.linspace(0, 1, 6) # start, end, num-points
>>> a.ndim >>> c
1 array([0. , 0.2, 0.4, 0.6, 0.8, 1. ])
>>> a.shape >>> d = np.linspace(0, 1, 5, endpoint=False)
(4,) >>> d
>>> len(a) array([0. , 0.2, 0.4, 0.6, 0.8])
4
• Common arrays:
• 2-D, 3-D, . . . :

3.1. The NumPy array object 58 3.1. The NumPy array object 59
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

>>> a = np.ones((3, 3)) # reminder: (3, 3) is a tuple Different data-types allow us to store data more compactly in memory, but most of the time we
>>> a simply work with floating point numbers. Note that, in the example above, NumPy auto-detects the
array([[1., 1., 1.], data-type from the input.
[1., 1., 1.],
[1., 1., 1.]])
>>> b = np.zeros((2, 2))
>>> b You can explicitly specify which data-type you want:
array([[0., 0.],
[0., 0.]]) >>> c = np.array([1, 2, 3], dtype=float)
>>> c = np.eye(3) >>> c.dtype
>>> c dtype('float64')
array([[1., 0., 0.],
[0., 1., 0.], The default data type is floating point:
[0., 0., 1.]])
>>> d = np.diag(np.array([1, 2, 3, 4])) >>> a = np.ones((3, 3))
>>> d >>> a.dtype
array([[1, 0, 0, 0], dtype('float64')
[0, 2, 0, 0],
[0, 0, 3, 0], There are also other types:
[0, 0, 0, 4]])
Complex
• np.random: random numbers (Mersenne Twister PRNG): >>> d = np.array([1+2j, 3+4j, 5+6*1j])
>>> d.dtype
>>> rng = np.random.default_rng(27446968) dtype('complex128')
>>> a = rng.random(4) # uniform in [0, 1]
>>> a
Bool
array([0.64613018, 0.48984931, 0.50851229, 0.22563948])
>>> e = np.array([True, False, False, True])
>>> b = rng.standard_normal(4) # Gaussian >>> e.dtype
>>> b dtype('bool')
array([-0.38250769, -0.61536465, 0.98131732, 0.59353096])
Strings

Exercise: Creating arrays using functions >>> f = np.array(['Bonjour', 'Hello', 'Hallo'])

>>> f.dtype # <--- strings containing max. 7 letters
• Experiment with arange, linspace, ones, zeros, eye and diag. dtype('<U7')
• Create different kinds of arrays with random numbers.
Much more
• Try setting the seed before creating an array with random values.
• int32
• Look at the function np.empty. What does it do? When might this be useful?
• int64
• uint32
3.1.3 Basic data types
• uint64
You may have noticed that, in some instances, array elements are displayed with a trailing dot (e.g. 2.
vs 2). This is due to a difference in the data-type used: 3.1.4 Basic visualization
>>> a = np.array([1, 2, 3]) Now that we have our first data arrays, we are going to visualize them.
>>> a.dtype
dtype('int64') Start by launching IPython:

$ ipython # or ipython3 depending on your install

>>> b = np.array([1., 2., 3.])
>>> b.dtype
Or the notebook:
dtype('float64')
$ jupyter notebook

 Tip Once IPython has started, enable interactive plots:

3.1. The NumPy array object 60 3.1. The NumPy array object 61
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

>>> %matplotlib

Or, from the notebook, enable plots in the notebook:

>>> %matplotlib inline

The inline is important for the notebook, so that plots are displayed in the notebook and not in a new
window.
Matplotlib is a 2D plotting package. We can import its functions as below:

>>> import matplotlib.pyplot as plt # the tidy way

And then use (note that you have to use show explicitly if you have not enabled interactive plots with
%matplotlib):

>>> plt.plot(x, y) # line plot

ã See also
>>> plt.show() # <-- shows the plot (not needed with interactive plots)
More in the: matplotlib chapter
Or, if you have enabled interactive plots with %matplotlib:

>>> plt.plot(x, y) # line plot

Exercise: Simple visualizations
• 1D plotting:
• Plot some simple arrays: a cosine as a function of time and a 2D matrix.
>>> x = np.linspace(0, 3, 20) • Try using the gray colormap on the 2D matrix.
>>> y = np.linspace(0, 9, 20)
>>> plt.plot(x, y) # line plot
[<matplotlib.lines.Line2D object at ...>] 3.1.5 Indexing and slicing
>>> plt.plot(x, y, 'o') # dot plot
[<matplotlib.lines.Line2D object at ...>] The items of an array can be accessed and assigned to the same way as other Python sequences (e.g.
lists):

>>> a = np.arange(10)
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> a[0], a[2], a[-1]
(np.int64(0), np.int64(2), np.int64(9))

. Warning

Indices begin at 0, like other Python sequences (and C/C++). In contrast, in Fortran or Matlab,
indices begin at 1.
• 2D arrays (such as images):
The usual python idiom for reversing a sequence is supported:
>>> rng = np.random.default_rng(27446968)
>>> image = rng.random((30, 30)) >>> a[::-1]
>>> plt.imshow(image, cmap=plt.cm.hot) array([9, 8, 7, 6, 5, 4, 3, 2, 1, 0])
<matplotlib.image.AxesImage object at ...>
>>> plt.colorbar() For multidimensional arrays, indices are tuples of integers:
<matplotlib.colorbar.Colorbar object at ...>
>>> a = np.diag(np.arange(3))
>>> a
array([[0, 0, 0],
[0, 1, 0],
[0, 0, 2]])
>>> a[1, 1]
(continues on next page)

3.1. The NumPy array object 62 3.1. The NumPy array object 63
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> a = np.arange(10)
np.int64(1) >>> a[5:] = 10
>>> a[2, 1] = 10 # third line, second column >>> a
>>> a array([ 0, 1, 2, 3, 4, 10, 10, 10, 10, 10])
array([[ 0, 0, 0], >>> b = np.arange(5)
[ 0, 1, 0], >>> a[5:] = b[::-1]
[ 0, 10, 2]]) >>> a
>>> a[1] array([0, 1, 2, 3, 4, 4, 3, 2, 1, 0])
array([0, 1, 0])

Exercise: Indexing and slicing

ò Note
• Try the different flavours of slicing, using start, end and step: starting from a linspace, try
• In 2D, the first dimension corresponds to rows, the second to columns. to obtain odd numbers counting backwards, and even numbers counting forwards.
• for multidimensional a, a[0] is interpreted by taking all elements in the unspecified dimensions. • Reproduce the slices in the diagram above. You may use the following expression to create the
array:
Slicing: Arrays, like other Python sequences can also be sliced: >>> np.arange(6) + np.arange(0, 51, 10)[:, np.newaxis]
array([[ 0, 1, 2, 3, 4, 5],
>>> a = np.arange(10) [10, 11, 12, 13, 14, 15],
>>> a [20, 21, 22, 23, 24, 25],
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]) [30, 31, 32, 33, 34, 35],
>>> a[2:9:3] # [start:end:step] [40, 41, 42, 43, 44, 45],
array([2, 5, 8]) [50, 51, 52, 53, 54, 55]])

Note that the last index is not included! :

>>> a[:4] Exercise: Array creation

array([0, 1, 2, 3])
Create the following arrays (with correct data types):
All three slice components are not required: by default, start is 0, end is the last and step is 1:
[[1, 1, 1, 1],
>>> a[1:3] [1, 1, 1, 1],
array([1, 2]) [1, 1, 1, 2],
>>> a[::2] [1, 6, 1, 1]]
array([0, 2, 4, 6, 8])
>>> a[3:] [[0., 0., 0., 0., 0.],
array([3, 4, 5, 6, 7, 8, 9]) [2., 0., 0., 0., 0.],
[0., 3., 0., 0., 0.],
A small illustrated summary of NumPy indexing and slicing. . . [0., 0., 4., 0., 0.],
[0., 0., 0., 5., 0.],
>>> a[0, 3:5] [0., 0., 0., 0., 6.]]
array([3, 4]) 0 1 2 3 4 5 Par on course: 3 statements for each
>>> a[4:, 4:] Hint: Individual array elements can be accessed similarly to a list, e.g. a[1] or a[1, 2].
10 11 12 13 14 15
array([[44, 55], Hint: Examine the docstring for diag.
[54, 55]]) 20 21 22 23 24 25
>>> a[:, 2] Exercise: Tiling for array creation
30 31 32 33 34 35
a([2, 12, 22, 32, 42, 52])
Skim through the documentation for np.tile, and use this function to construct the array:
>>> a[2::2, ::2] 40 41 42 43 44 45 [[4, 3, 4, 3, 4, 3],
array([[20, 22, 24], [2, 1, 2, 1, 2, 1],
50 51 52 53 54 55 [4, 3, 4, 3, 4, 3],
[40, 42, 44]]) [2, 1, 2, 1, 2, 1]]
You can also combine assignment and slicing:

3.1. The NumPy array object 64 3.1. The NumPy array object 65
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

3.1.6 Copies and views • Cross out 0 and 1 which are not primes:
A slicing operation creates a view on the original array, which is just a way of accessing array data. >>> is_prime[:2] = 0
Thus the original array is not copied in memory. You can use np.may_share_memory() to check if two
arrays share the same memory block. Note however, that this uses heuristics and may give you false • For each integer j starting from 2, cross out its higher multiples:
positives.
>>> N_max = int(np.sqrt(len(is_prime) - 1))
When modifying the view, the original array is modified as well: >>> for j in range(2, N_max + 1):
... is_prime[2*j::j] = False
>>> a = np.arange(10)
>>> a • Skim through help(np.nonzero), and print the prime numbers
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
• Follow-up:
>>> b = a[::2]
>>> b – Move the above code into a script file named prime_sieve.py
array([0, 2, 4, 6, 8])
– Run it to check it works
>>> np.may_share_memory(a, b)
True – Use the optimization suggested in the sieve of Eratosthenes:
>>> b[0] = 12
1. Skip j which are already known to not be primes
>>> b
array([12, 2, 4, 6, 8]) 2. The first number to cross out is 𝑗 2
>>> a # (!)
array([12, 1, 2, 3, 4, 5, 6, 7, 8, 9])
3.1.7 Fancy indexing
>>> a = np.arange(10)
>>> c = a[::2].copy() # force a copy
>>> c[0] = 12  Tip
>>> a
NumPy arrays can be indexed with slices, but also with boolean or integer arrays (masks). This
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
method is called fancy indexing. It creates copies not views.
>>> np.may_share_memory(a, c)
False Using boolean masks

This behavior can be surprising at first sight. . . but it allows to save both memory and time. >>> rng = np.random.default_rng(27446968)
>>> a = rng.integers(0, 21, 15)
>>> a
Worked example: Prime number sieve array([ 3, 13, 12, 10, 10, 10, 18, 4, 8, 5, 6, 11, 12, 17, 3])
>>> (a % 3 == 0)
array([ True, False, True, False, False, False, True, False, False,
False, True, False, True, False, True])
>>> mask = (a % 3 == 0)
>>> extract_from_a = a[mask] # or, a[a%3==0]
>>> extract_from_a # extract a sub-array with the mask
array([ 3, 12, 18, 6, 12, 3])

Indexing with a mask can be very useful to assign a new value to a sub-array:

>>> a[a % 3 == 0] = -1
>>> a
array([-1, 13, -1, 10, 10, 10, -1, 4, 8, 5, -1, 11, -1, 17, -1])

Indexing with an array of integers

>>> a = np.arange(0, 100, 10)

>>> a
Compute prime numbers in 0–99, with a sieve array([ 0, 10, 20, 30, 40, 50, 60, 70, 80, 90])

• Construct a shape (100,) boolean array is_prime, filled with True in the beginning: Indexing can be done with an array of integers, where the same index is repeated several time:
>>> is_prime = np.ones((100,), dtype=bool)

3.1. The NumPy array object 66 3.1. The NumPy array object 67
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

>>> a[[2, 3, 2, 4, 2]] # note: [2, 3, 2, 4, 2] is a Python list • Elementwise operations

array([20, 30, 20, 40, 20])
• Basic reductions
New values can be assigned with this kind of indexing: • Broadcasting

>>> a[[9, 7]] = -100 • Array shape manipulation

>>> a • Sorting data
array([ 0, 10, 20, 30, 40, 50, 60, -100, 80, -100])
• Summary

 Tip

When a new array is created by indexing with an array of integers, the new array has the same shape 3.2.1 Elementwise operations
as the array of integers: Basic operations
>>> a = np.arange(10) With scalars:
>>> idx = np.array([[3, 4], [9, 7]])
>>> idx.shape >>> a = np.array([1, 2, 3, 4])
(2, 2) >>> a + 1
>>> a[idx] array([2, 3, 4, 5])
array([[3, 4], >>> 2**a
[9, 7]]) array([ 2, 4, 8, 16])

All arithmetic operates elementwise:

>>> b = np.ones(4) + 1
The image below illustrates various fancy indexing applications >>> a - b
array([-1., 0., 1., 2.])
>>> a[(0,1,2,3,4), (1,2,3,4,5)] >>> a * b
0 1 2 3 4 5 array([2., 4., 6., 8.])
array([1, 12, 23, 34, 45])

>>> a[3:, [0,2,5]] 10 11 12 13 14 15 >>> j = np.arange(5)

array([[30, 32, 35], >>> 2**(j + 1) - j
20 21 22 23 24 25 array([ 2, 3, 6, 13, 28])
[40, 42, 45],
[50, 52, 55]])
30 31 32 33 34 35 These operations are of course much faster than if you did them in pure python:
>>> mask = np.array([1,0,1,0,0,1], dtype=bool) >>> a = np.arange(10000)
>>> a[mask, 2] 40 41 42 43 44 45
>>> %timeit a + 1
array([2, 22, 52]) 10000 loops, best of 3: 24.3 us per loop
50 51 52 53 54 55 >>> l = range(10000)
>>> %timeit [i+1 for i in l]
1000 loops, best of 3: 861 us per loop
Exercise: Fancy indexing

• Again, reproduce the fancy indexing shown in the diagram above.

. Warning
• Use fancy indexing on the left and array creation on the right to assign values into an array,
for instance by setting parts of the array in the diagram above to zero. Array multiplication is not matrix multiplication:
>>> c = np.ones((3, 3))
>>> c * c # NOT matrix multiplication!
array([[1., 1., 1.],
[1., 1., 1.],
3.2 Numerical operations on arrays [1., 1., 1.]])

Section contents
ò Note

3.2. Numerical operations on arrays 68 3.2. Numerical operations on arrays 69

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

Matrix multiplication:
array([ 1. , 2.71828183, 7.3890561 , 20.08553692, 54.59815003])
>>> c @ c
>>> np.log(np.exp(a))
array([[3., 3., 3.],
array([0., 1., 2., 3., 4.])
[3., 3., 3.],
[3., 3., 3.]])
Shape mismatches

>>> a = np.arange(4)
Exercise: Elementwise operations >>> a + np.array([1, 2])
Traceback (most recent call last):
• Try simple arithmetic elementwise operations: add even elements with odd elements File "<stdin>", line 1, in <module>
• Time them against their pure python counterparts using %timeit. ValueError: operands could not be broadcast together with shapes (4,) (2,)

• Generate: Broadcasting? We’ll return to that later.

– [2**0, 2**1, 2**2, 2**3, 2**4] Transposition:
– a_j = 2^(3*j) - j
>>> a = np.triu(np.ones((3, 3)), 1) # see help(np.triu)
>>> a
array([[0., 1., 1.],
Other operations
[0., 0., 1.],
Comparisons: [0., 0., 0.]])
>>> a.T
>>> a = np.array([1, 2, 3, 4]) array([[0., 0., 0.],
>>> b = np.array([4, 2, 2, 4]) [1., 0., 0.],
>>> a == b [1., 1., 0.]])
array([False, True, False, True])
>>> a > b
array([False, False, True, False])
ò Note

The transposition is a view

 Tip
The transpose returns a view of the original array:
Array-wise comparisons: >>> a = np.arange(9).reshape(3, 3)
>>> a = np.array([1, 2, 3, 4]) >>> a.T[0, 2] = 999
>>> b = np.array([4, 2, 2, 4]) >>> a.T
>>> c = np.array([1, 2, 3, 4]) array([[ 0, 3, 999],
>>> np.array_equal(a, b) [ 1, 4, 7],
False [ 2, 5, 8]])
>>> np.array_equal(a, c) >>> a
True array([[ 0, 1, 2],
[ 3, 4, 5],
[999, 7, 8]])
Logical operations:

>>> a = np.array([1, 1, 0, 0], dtype=bool)

>>> b = np.array([1, 0, 1, 0], dtype=bool) ò Note
>>> np.logical_or(a, b)
array([ True, True, True, False]) Linear algebra
>>> np.logical_and(a, b)
The sub-module numpy.linalg implements basic linear algebra, such as solving linear systems, sin-
array([ True, False, False, False])
gular value decomposition, etc. However, it is not guaranteed to be compiled using efficient routines,
and thus we recommend the use of scipy.linalg, as detailed in section Linear algebra operations:
Transcendental functions: scipy.linalg
>>> a = np.arange(5)
>>> np.sin(a)
array([ 0. , 0.84147098, 0.90929743, 0.14112001, -0.7568025 ]) Exercise other operations
>>> np.exp(a)
(continues on next page)
• Look at the help for np.allclose. When might this be useful?

3.2. Numerical operations on arrays 70 3.2. Numerical operations on arrays 71

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• Look at the help for np.triu and np.tril. Other reductions

— works the same way (and take axis=)

3.2.2 Basic reductions Extrema:

Computing sums >>> x = np.array([1, 3, 2])

>>> x.min()
>>> x = np.array([1, 2, 3, 4]) np.int64(1)
>>> np.sum(x) >>> x.max()
np.int64(10) np.int64(3)
>>> x.sum()
np.int64(10) >>> x.argmin() # index of minimum
np.int64(0)
>>> x.argmax() # index of maximum
np.int64(1)

Logical operations:

>>> np.all([True, True, False])

np.False_
>>> np.any([True, True, False])
np.True_

ò Note

Can be used for array comparisons:

>>> a = np.zeros((100, 100))
>>> np.any(a != 0)
Sum by rows and by columns: np.False_
>>> np.all(a == a)
>>> x = np.array([[1, 1], [2, 2]])
np.True_
>>> x
array([[1, 1],
>>> a = np.array([1, 2, 3, 2])
[2, 2]])
>>> b = np.array([2, 2, 3, 2])
>>> x.sum(axis=0) # columns (first dimension)
>>> c = np.array([6, 4, 4, 5])
array([3, 3])
>>> ((a <= b) & (b <= c)).all()
>>> x[:, 0].sum(), x[:, 1].sum()
np.True_
(np.int64(3), np.int64(3))
>>> x.sum(axis=1) # rows (second dimension)
array([2, 4])
>>> x[0, :].sum(), x[1, :].sum() Statistics:
(np.int64(2), np.int64(4)) >>> x = np.array([1, 2, 3, 1])
>>> y = np.array([[1, 2, 3], [5, 6, 1]])
>>> x.mean()
 Tip np.float64(1.75)
>>> np.median(x)
Same idea in higher dimensions: np.float64(1.5)
>>> rng = np.random.default_rng(27446968) >>> np.median(y, axis=-1) # last axis
>>> x = rng.random((2, 2, 2)) array([2., 5.])
>>> x.sum(axis=2)[0, 1]
np.float64(0.73415...) >>> x.std() # full population standard dev.
>>> x[0, 1, :].sum() np.float64(0.82915619758884995)
np.float64(0.73415...)
. . . and many more (best to learn as you go).

3.2. Numerical operations on arrays 72 3.2. Numerical operations on arrays 73

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Exercise: Reductions >>> plt.figure(figsize=(4, 3))

• Given there is a sum, what other function might you expect to see? <Figure size ... with 0 Axes>
>>> plt.plot(t, np.sqrt(mean_sq_distance), 'g.', t, np.sqrt(t), 'y-')
• What is the difference between sum and cumsum? [<matplotlib.lines.Line2D object at ...>, <matplotlib.lines.Line2D object at ...>]
>>> plt.xlabel(r"$t$")
Text(...'$t$')
Worked Example: diffusion using a random walk algorithm >>> plt.ylabel(r"$\sqrt{\langle (\delta x)^2 \rangle}$")
Text(...'$\\sqrt{\\langle (\\delta x)^2 \\rangle}$')
>>> plt.tight_layout() # provide sufficient space for labels

 Tip

Let us consider a simple 1D random walk process: at each time step a walker jumps right or left
with equal probability.
We are interested in finding the typical distance from the origin of a random walker after t left
or right jumps? We are going to simulate many “walkers” to find this law, and we are going to
do so using array computing tricks: we are going to create a 2D array with the “stories” (each
walker has a story) in one direction, and the time in the other:

We find a well-known result in physics: the RMS distance grows as the square root of the time!

3.2.3 Broadcasting
• Basic operations on numpy arrays (addition, etc.) are elementwise
• This works on arrays of the same size.
Nevertheless, It’s also possible to do operations on arrays of different
sizes if NumPy can transform these arrays so that they all have
the same size: this conversion is called broadcasting.
>>> n_stories = 1000 # number of walkers The image below gives an example of broadcasting:
>>> t_max = 200 # time during which we follow the walker

We randomly choose all the steps 1 or -1 of the walk:

>>> t = np.arange(t_max)
>>> rng = np.random.default_rng()
>>> steps = 2 * rng.integers(0, 1 + 1, (n_stories, t_max)) - 1 # +1 because the␣
˓→high value is exclusive

>>> np.unique(steps) # Verification: all steps are 1 or -1

array([-1, 1])

We build the walks by summing steps along the time:

>>> positions = np.cumsum(steps, axis=1) # axis = 1: dimension of time
>>> sq_distance = positions**2

We get the mean in the axis of the stories:

>>> mean_sq_distance = np.mean(sq_distance, axis=0)

Plot the results:

3.2. Numerical operations on arrays 74 3.2. Numerical operations on arrays 75

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> a = a[:, np.newaxis] # adds a new axis -> 2D array
>>> a.shape
(4, 1)
>>> a
array([[ 0],
[10],
[20],
[30]])
>>> a + b
array([[ 0, 1, 2],
[10, 11, 12],
[20, 21, 22],
[30, 31, 32]])

 Tip

Broadcasting seems a bit magical, but it is actually quite natural to use it when we want to solve a
problem whose output data is an array with more dimensions than input data.

Worked Example: Broadcasting

Let’s construct an array of distances (in miles) between cities of Route 66: Chicago, Springfield,
Saint-Louis, Tulsa, Oklahoma City, Amarillo, Santa Fe, Albuquerque, Flagstaff and Los Angeles.
>>> mileposts = np.array([0, 198, 303, 736, 871, 1175, 1475, 1544,
Let’s verify: ... 1913, 2448])
>>> distance_array = np.abs(mileposts - mileposts[:, np.newaxis])
>>> a = np.tile(np.arange(0, 40, 10), (3, 1)).T >>> distance_array
>>> a array([[ 0, 198, 303, 736, 871, 1175, 1475, 1544, 1913, 2448],
array([[ 0, 0, 0], [ 198, 0, 105, 538, 673, 977, 1277, 1346, 1715, 2250],
[10, 10, 10], [ 303, 105, 0, 433, 568, 872, 1172, 1241, 1610, 2145],
[20, 20, 20], [ 736, 538, 433, 0, 135, 439, 739, 808, 1177, 1712],
[30, 30, 30]]) [ 871, 673, 568, 135, 0, 304, 604, 673, 1042, 1577],
>>> b = np.array([0, 1, 2]) [1175, 977, 872, 439, 304, 0, 300, 369, 738, 1273],
>>> a + b [1475, 1277, 1172, 739, 604, 300, 0, 69, 438, 973],
array([[ 0, 1, 2], [1544, 1346, 1241, 808, 673, 369, 69, 0, 369, 904],
[10, 11, 12], [1913, 1715, 1610, 1177, 1042, 738, 438, 369, 0, 535],
[20, 21, 22], [2448, 2250, 2145, 1712, 1577, 1273, 973, 904, 535, 0]])
[30, 31, 32]])

We have already used broadcasting without knowing it!:

>>> a = np.ones((4, 5))

>>> a[0] = 2 # we assign an array of dimension 0 to an array of dimension 1
>>> a
array([[2., 2., 2., 2., 2.],
[1., 1., 1., 1., 1.],
[1., 1., 1., 1., 1.],
[1., 1., 1., 1., 1.]])

A useful trick:

>>> a = np.arange(0, 40, 10)

>>> a.shape
(4,)
(continues on next page)

3.2. Numerical operations on arrays 76 3.2. Numerical operations on arrays 77

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

(array([[0],
[1],
[2],
[3],
[4]]), array([[0, 1, 2, 3, 4]]))
>>> x.shape, y.shape
((5, 1), (1, 5))
>>> distance = np.sqrt(x ** 2 + y ** 2)

 Tip

So, np.ogrid is very useful as soon as we have to handle computations on a grid. On the other hand,
np.mgrid directly provides matrices full of indices for cases where we can’t (or don’t want to) benefit
from broadcasting:
>>> x, y = np.mgrid[0:4, 0:4]
>>> x
array([[0, 0, 0, 0],
A lot of grid-based or network-based problems can also use broadcasting. For instance, if we want to [1, 1, 1, 1],
compute the distance from the origin of points on a 5x5 grid, we can do [2, 2, 2, 2],
>>> x, y = np.arange(5), np.arange(5)[:, np.newaxis] [3, 3, 3, 3]])
>>> distance = np.sqrt(x ** 2 + y ** 2) >>> y
>>> distance array([[0, 1, 2, 3],
array([[0. , 1. , 2. , 3. , 4. ], [0, 1, 2, 3],
[1. , 1.41421356, 2.23606798, 3.16227766, 4.12310563], [0, 1, 2, 3],
[2. , 2.23606798, 2.82842712, 3.60555128, 4.47213595], [0, 1, 2, 3]])
[3. , 3.16227766, 3.60555128, 4.24264069, 5. ],
[4. , 4.12310563, 4.47213595, 5. , 5.65685425]])

Or in color: ã See also

>>> plt.pcolor(distance) Broadcasting: discussion of broadcasting in the Advanced NumPy chapter.

<matplotlib.collections.PolyQuadMesh object at ...>
>>> plt.colorbar()
<matplotlib.colorbar.Colorbar object at ...> 3.2.4 Array shape manipulation
Flattening

>>> a = np.array([[1, 2, 3], [4, 5, 6]])

>>> a.ravel()
array([1, 2, 3, 4, 5, 6])
>>> a.T
array([[1, 4],
[2, 5],
[3, 6]])
>>> a.T.ravel()
array([1, 4, 2, 5, 3, 6])

Higher dimensions: last dimensions ravel out “first”.

Reshaping
Remark : the numpy.ogrid() function allows to directly create vectors x and y of the previous example, The inverse operation to flattening:
with two “significant dimensions”:
>>> a.shape
>>> x, y = np.ogrid[0:5, 0:5] (2, 3)
>>> x, y >>> b = a.ravel()
(continues on next page) (continues on next page)

3.2. Numerical operations on arrays 78 3.2. Numerical operations on arrays 79

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) Dimension shuffling

>>> b = b.reshape((2, 3))
>>> a = np.arange(4*3*2).reshape(4, 3, 2)
>>> b
>>> a.shape
array([[1, 2, 3],
(4, 3, 2)
[4, 5, 6]])
>>> a[0, 2, 1]
np.int64(5)
Or,
>>> b = a.transpose(1, 2, 0)
>>> a.reshape((2, -1)) # unspecified (-1) value is inferred >>> b.shape
array([[1, 2, 3], (3, 2, 4)
[4, 5, 6]]) >>> b[2, 1, 0]
np.int64(5)

Also creates a view:

. Warning
>>> b[2, 1, 0] = -1
ndarray.reshape may return a view (cf help(np.reshape))), or copy
>>> a[0, 2, 1]
np.int64(-1)

 Tip Resizing
>>> b[0, 0] = 99 Size of an array can be changed with ndarray.resize:
>>> a
array([[99, 2, 3], >>> a = np.arange(4)
[ 4, 5, 6]]) >>> a.resize((8,))
>>> a
Beware: reshape may also return a copy!: array([0, 1, 2, 3, 0, 0, 0, 0])
>>> a = np.zeros((3, 2))
>>> b = a.T.reshape(3*2) However, it must not be referred to somewhere else:
>>> b[0] = 9 >>> b = a
>>> a >>> a.resize((4,))
array([[0., 0.], Traceback (most recent call last):
[0., 0.], File "<stdin>", line 1, in <module>
[0., 0.]]) ValueError: cannot resize an array that references or is referenced
To understand this you need to learn more about the memory layout of a NumPy array. by another array in this way.
Use the np.resize function or refcheck=False

Adding a dimension
Exercise: Shape manipulations
Indexing with the np.newaxis object allows us to add an axis to an array (you have seen this already
above in the broadcasting section): • Look at the docstring for reshape, especially the notes section which has some more information
about copies and views.
>>> z = np.array([1, 2, 3])
>>> z • Use flatten as an alternative to ravel. What is the difference? (Hint: check which one returns
array([1, 2, 3]) a view and which a copy)
• Experiment with transpose for dimension shuffling.
>>> z[:, np.newaxis]
array([[1],
[2],
3.2.5 Sorting data
[3]])
Sorting along an axis:
>>> z[np.newaxis, :]
array([[1, 2, 3]]) >>> a = np.array([[4, 3, 5], [1, 2, 1]])
>>> b = np.sort(a, axis=1)
>>> b
array([[3, 4, 5],
[1, 1, 2]])

3.2. Numerical operations on arrays 80 3.2. Numerical operations on arrays 81

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

ò Note
Quick read
Sorts each row separately!
If you want to do a first quick pass through the Scientific Python Lectures to learn the ecosystem,
you can directly skip to the next chapter: Matplotlib: plotting.
In-place sort: The remainder of this chapter is not necessary to follow the rest of the intro part. But be sure to
>>> a.sort(axis=1) come back and finish this chapter, as well as to do some more exercises.
>>> a
array([[3, 4, 5],
[1, 1, 2]])

Sorting with fancy indexing:

3.3 More elaborate arrays
>>> a = np.array([4, 3, 1, 2])
>>> j = np.argsort(a) Section contents
>>> j
array([2, 3, 1, 0]) • More data types
>>> a[j] • Structured data types
array([1, 2, 3, 4])
• maskedarray: dealing with (propagation of) missing data
Finding minima and maxima:

>>> a = np.array([4, 3, 1, 2])

>>> j_max = np.argmax(a) 3.3.1 More data types
>>> j_min = np.argmin(a)
>>> j_max, j_min Casting
(np.int64(0), np.int64(2)) “Bigger” type wins in mixed-type operations:

>>> np.array([1, 2, 3]) + 1.5

Exercise: Sorting array([2.5, 3.5, 4.5])

• Try both in-place and out-of-place sorting. Assignment never changes the type!
• Try creating arrays with different dtypes and sorting them. >>> a = np.array([1, 2, 3])
• Use all or array_equal to check the results. >>> a.dtype
dtype('int64')
• Look at np.random.shuffle for a way to create sortable input quicker. >>> a[0] = 1.9 # <-- float is truncated to integer
• Combine ravel, sort and reshape. >>> a
array([1, 2, 3])
• Look at the axis keyword for sort and rewrite the previous exercise.
Forced casts:

3.2.6 Summary >>> a = np.array([1.7, 1.2, 1.6])

>>> b = a.astype(int) # <-- truncates to integer
What do you need to know to get started?
>>> b
• Know how to create arrays : array, arange, ones, zeros. array([1, 1, 1])
• Know the shape of the array with array.shape, then use slicing to obtain different views of the
Rounding:
array: array[::2], etc. Adjust the shape of the array using reshape or flatten it with ravel.
• Obtain a subset of the elements of an array and/or modify their values with masks >>> a = np.array([1.2, 1.5, 1.6, 2.5, 3.5, 4.5])
>>> b = np.around(a)
>>> a[a < 0] = 0 >>> b # still floating-point
array([1., 2., 2., 2., 4., 4.])
• Know miscellaneous operations on arrays, such as finding the mean or max (array.max(), array. >>> c = np.around(a).astype(int)
mean()). No need to retain everything, but have the reflex to search in the documentation (online >>> c
docs, help())!! array([1, 2, 2, 2, 4, 4])
• For advanced use: master the indexing with arrays of integers, as well as broadcasting. Know more
NumPy functions to handle various array operations.

3.2. Numerical operations on arrays 82 3.3. More elaborate arrays 83

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Different data type sizes • Half the size in memory and on disk
Integers (signed): • Half the memory bandwidth required (may be a bit faster in some operations)
In [1]: a = np.zeros((int(1e6),), dtype=np.float64)
int8 8 bits
int16 16 bits In [2]: b = np.zeros((int(1e6),), dtype=np.float32)
int32 32 bits (same as int on 32-bit platform)
int64 64 bits (same as int on 64-bit platform) In [3]: %timeit a*a
269 us +- 2.39 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)

>>> np.array([1], dtype=int).dtype In [4]: %timeit b*b

dtype('int64') 135 us +- 135 ns per loop (mean +- std. dev. of 7 runs, 10,000 loops each)
>>> np.iinfo(np.int32).max, 2**31 - 1
(2147483647, 2147483647) • But: bigger rounding errors — sometimes in surprising places (i.e., don’t use them unless you
really need them)
Unsigned integers:

3.3.2 Structured data types

uint8 8 bits
uint16 16 bits
uint32 32 bits sensor_code (4-character string)
uint64 64 bits position (float)
value (float)

>>> np.iinfo(np.uint32).max, 2**32 - 1

(4294967295, 4294967295) >>> samples = np.zeros((6,), dtype=[('sensor_code', 'S4'),
... ('position', float), ('value', float)])
Floating-point numbers: >>> samples.ndim
1
>>> samples.shape
float16 16 bits (6,)
float32 32 bits >>> samples.dtype.names
float64 64 bits (same as float) ('sensor_code', 'position', 'value')
float96 96 bits, platform-dependent (same as np.longdouble) >>> samples[:] = [('ALFA', 1, 0.37), ('BETA', 1, 0.11), ('TAU', 1, 0.13),
float128 128 bits, platform-dependent (same as np.longdouble) ... ('ALFA', 1.5, 0.37), ('ALFA', 3, 0.11), ('TAU', 1.2, 0.13)]
>>> samples
>>> np.finfo(np.float32).eps array([(b'ALFA', 1. , 0.37), (b'BETA', 1. , 0.11), (b'TAU', 1. , 0.13),
np.float32(1.1920929e-07) (b'ALFA', 1.5, 0.37), (b'ALFA', 3. , 0.11), (b'TAU', 1.2, 0.13)],
>>> np.finfo(np.float64).eps dtype=[('sensor_code', 'S4'), ('position', '<f8'), ('value', '<f8')])
np.float64(2.220446049250313e-16)
Field access works by indexing with field names:
>>> np.float32(1e-8) + np.float32(1) == 1
>>> samples['sensor_code']
np.True_
array([b'ALFA', b'BETA', b'TAU', b'ALFA', b'ALFA', b'TAU'], dtype='|S4')
>>> np.float64(1e-8) + np.float64(1) == 1
>>> samples['value']
np.False_
array([0.37, 0.11, 0.13, 0.37, 0.11, 0.13])
>>> samples[0]
Complex floating-point numbers:
np.void((b'ALFA', 1.0, 0.37), dtype=[('sensor_code', 'S4'), ('position', '<f8'), (
˓→'value', '<f8')])

complex64 two 32-bit floats

complex128 two 64-bit floats >>> samples[0]['sensor_code'] = 'TAU'
complex192 two 96-bit floats, platform-dependent >>> samples[0]
complex256 two 128-bit floats, platform-dependent np.void((b'TAU', 1.0, 0.37), dtype=[('sensor_code', 'S4'), ('position', '<f8'), (
˓→'value', '<f8')])

Smaller data types Multiple fields at once:

If you don’t know you need special data types, then you probably don’t. >>> samples[['position', 'value']]
array([(1. , 0.37), (1. , 0.11), (1. , 0.13), (1.5, 0.37),
Comparison on using float32 instead of float64:
(continues on next page)

3.3. More elaborate arrays 84 3.3. More elaborate arrays 85

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) • Except some rare cases, variable names and comments in English.
(3. , 0.11), (1.2, 0.13)],
dtype={'names': ['position', 'value'], 'formats': ['<f8', '<f8'], 'offsets': [4,
˓→ 12], 'itemsize': 20})

Fancy indexing works, as usual: 3.4 Advanced operations

>>> samples[samples['sensor_code'] == b'ALFA']
array([(b'ALFA', 1.5, 0.37), (b'ALFA', 3. , 0.11)], Section contents
dtype=[('sensor_code', 'S4'), ('position', '<f8'), ('value', '<f8')])
• Polynomials
• Loading data files
ò Note

There are a bunch of other syntaxes for constructing structured arrays, see here and here.

3.4.1 Polynomials
3.3.3 maskedarray: dealing with (propagation of) missing data NumPy also contains polynomials in different bases:
• For floats one could use NaN’s, but masks work for all types: For example, 3𝑥2 + 2𝑥 − 1:
>>> x = np.ma.array([1, 2, 3, 4], mask=[0, 1, 0, 1]) >>> p = np.poly1d([3, 2, -1])
>>> x >>> p(0)
masked_array(data=[1, --, 3, --], np.int64(-1)
mask=[False, True, False, True], >>> p.roots
fill_value=999999) array([-1. , 0.33333333])
>>> p.order
2
>>> y = np.ma.array([1, 2, 3, 4], mask=[0, 1, 1, 1])
>>> x + y
>>> x = np.linspace(0, 1, 20)
masked_array(data=[2, --, --, --],
>>> rng = np.random.default_rng()
mask=[False, True, True, True],
>>> y = np.cos(x) + 0.3*rng.random(20)
fill_value=999999)
>>> p = np.poly1d(np.polyfit(x, y, 3))
• Masking versions of common functions:
>>> t = np.linspace(0, 1, 200) # use a larger number of points for smoother plotting
>>> np.ma.sqrt([1, -1, 2, -2]) >>> plt.plot(x, y, 'o', t, p(t), '-')
masked_array(data=[1.0, --, 1.41421356237... --], [<matplotlib.lines.Line2D object at ...>, <matplotlib.lines.Line2D object at ...>]
mask=[False, True, False, True],
fill_value=1e+20)

ò Note

There are other useful array siblings

While it is off topic in a chapter on NumPy, let’s take a moment to recall good coding practice, which
really do pay off in the long run:

Good practices

• Explicit variable names (no need of a comment to explain what is in the variable)
See https://numpy.org/doc/stable/reference/routines.polynomials.poly1d.html for more.
• Style: spaces after commas, around =, etc.
A certain number of rules for writing “beautiful” code (and, more importantly, using the same
conventions as everybody else!) are given in the Style Guide for Python Code and the Docstring
Conventions page (to manage help strings).

3.3. More elaborate arrays 86 3.4. Advanced operations 87

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

More polynomials (with more bases) (continued from previous page)

NumPy also has a more sophisticated polynomial interface, which supports e.g. the Chebyshev basis. [ 1902., 70200., 9800., 41500.],
...
3𝑥2 + 2𝑥 − 1:

>>> p = np.polynomial.Polynomial([-1, 2, 3]) # coefs in different order! >>> np.savetxt('pop2.txt', data)

>>> p(0) >>> data2 = np.loadtxt('pop2.txt')
np.float64(-1.0)
>>> p.roots()
array([-1. , 0.33333333]) ò Note
>>> p.degree() # In general polynomials do not always expose 'order'
2 If you have a complicated text file, what you can try are:
• np.genfromtxt
Example using polynomials in Chebyshev basis, for polynomials in range [-1, 1]:
• Using Python’s I/O functions and e.g. regexps for parsing (Python is quite well suited for this)
>>> x = np.linspace(-1, 1, 2000)
>>> rng = np.random.default_rng()
>>> y = np.cos(x) + 0.3*rng.random(2000) Reminder: Navigating the filesystem with IPython
>>> p = np.polynomial.Chebyshev.fit(x, y, 90)
In [1]: pwd # show current directory
>>> plt.plot(x, y, 'r.') Out[1]: '/tmp'
[<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(x, p(x), 'k-', lw=3)
[<matplotlib.lines.Line2D object at ...>]
Images
Using Matplotlib:

>>> img = plt.imread('data/elephant.png')

>>> img.shape, img.dtype
((200, 300, 3), dtype('float32'))
>>> plt.imshow(img)
<matplotlib.image.AxesImage object at ...>
>>> plt.savefig('plot.png')

>>> plt.imsave('red_elephant.png', img[:,:,0], cmap=plt.cm.gray)

The Chebyshev polynomials have some advantages in interpolation.

3.4.2 Loading data files

Text files
Example: populations.txt:
# year hare lynx carrot
1900 30e3 4e3 48300
1901 47.2e3 6.1e3 48200
1902 70.2e3 9.8e3 41500
1903 77.4e3 35.2e3 38200 This saved only one channel (of RGB):

>>> plt.imshow(plt.imread('red_elephant.png'))
>>> data = np.loadtxt('data/populations.txt') <matplotlib.image.AxesImage object at ...>
>>> data
array([[ 1900., 30000., 4000., 48300.],
[ 1901., 47200., 6100., 48200.],
(continues on next page)

3.4. Advanced operations 88 3.4. Advanced operations 89

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Exercise: Text data files

Write a Python script that loads data from populations.txt:: and drop the last column and the
first 5 rows. Save the smaller dataset to pop2.txt.

NumPy internals

If you are interested in the NumPy internals, there is a good discussion in Advanced NumPy.

3.5 Some exercises

Other libraries:
3.5.1 Array manipulations
>>> import imageio.v3 as iio 1. Form the 2-D array (without typing it in explicitly):
>>> iio.imwrite('tiny_elephant.png', (img[::6,::6] * 255).astype(np.uint8))
>>> plt.imshow(plt.imread('tiny_elephant.png'), interpolation='nearest') [[1, 6, 11],
<matplotlib.image.AxesImage object at ...> [2, 7, 12],
[3, 8, 13],
[4, 9, 14],
[5, 10, 15]]

and generate a new array containing its 2nd and 4th rows.
2. Divide each column of the array:

>>> import numpy as np

>>> a = np.arange(25).reshape(5, 5)

elementwise with the array b = np.array([1., 5, 10, 15, 20]). (Hint: np.newaxis).
3. Harder one: Generate a 10 x 3 array of random numbers (in range [0,1]). For each row, pick the
number closest to 0.5.
• Use abs and argmin to find the column j closest for each row.
• Use fancy indexing to extract the numbers. (Hint: a[i,j] – the array i must contain the
NumPy’s own format row numbers corresponding to stuff in j.)
NumPy has its own binary format, not portable but with efficient I/O:
3.5.2 Picture manipulation: Framing a Face
>>> data = np.ones((3, 3))
>>> np.save('pop.npy', data) Let’s do some manipulations on NumPy arrays by starting with an image of a raccoon. scipy provides
>>> data3 = np.load('pop.npy') a 2D array of this image with the scipy.datasets.face function:

>>> import scipy as sp

Well-known (& more obscure) file formats >>> face = sp.datasets.face(gray=True) # 2D grayscale image
• HDF5: h5py, PyTables
Here are a few images we will be able to obtain with our manipulations: use different colormaps, crop
• NetCDF: scipy.io.netcdf_file, netcdf4-python, . . . the image, change some parts of the image.
• Matlab: scipy.io.loadmat, scipy.io.savemat
• MatrixMarket: scipy.io.mmread, scipy.io.mmwrite
• IDL: scipy.io.readsav
. . . if somebody uses it, there’s probably also a Python library for it.

• Let’s use the imshow function of matplotlib to display the image.

3.4. Advanced operations 90 3.5. Some exercises 91

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

>>> import matplotlib.pyplot as plt

>>> face = sp.datasets.face(gray=True)
>>> plt.imshow(face)
<matplotlib.image.AxesImage object at 0x...>

• The face is displayed in false colors. A colormap must be

specified for it to be displayed in grey.

>>> plt.imshow(face, cmap=plt.cm.gray)

<matplotlib.image.AxesImage object at 0x...>

• Create an array of the image with a narrower centering

[for example,] remove 100 pixels from all the borders of the image. To check the result, display
this new array with imshow.

>>> crop_face = face[100:-100, 100:-100] Computes and print, based on the data in populations.txt. . .
1. The mean and std of the populations of each species for the years in the period.
• We will now frame the face with a black locket. For this, we
need to create a mask corresponding to the pixels we want to be black. The center of the face 2. Which year each species had the largest population.
is around (660, 330), so we defined the mask by this condition (y-300)**2 + (x-660)**2 3. Which species has the largest population for each year. (Hint: argsort & fancy indexing of
np.array(['H', 'L', 'C']))
>>> sy, sx = face.shape
>>> y, x = np.ogrid[0:sy, 0:sx] # x and y indices of pixels 4. Which years any of the populations is above 50000. (Hint: comparisons and np.any)
>>> y.shape, x.shape 5. The top 2 years for each species when they had the lowest populations. (Hint: argsort, fancy
((768, 1), (1, 1024)) indexing)
>>> centerx, centery = (660, 300) # center of the image
>>> mask = ((y - centery)**2 + (x - centerx)**2) > 230**2 # circle 6. Compare (plot) the change in hare population (see help(np.gradient)) and the number of lynxes.
Check correlation (see help(np.corrcoef)).
then we assign the value 0 to the pixels of the image corresponding to the mask. The syntax . . . all without for-loops.
is extremely simple and intuitive:
Solution: Python source file
>>> face[mask] = 0
>>> plt.imshow(face) 3.5.4 Crude integral approximations
<matplotlib.image.AxesImage object at 0x...>
Write a function f(a, b, c) that returns 𝑎𝑏 −𝑐. Form a 24x12x6 array containing its values in parameter
• Follow-up: copy all instructions of this exercise in a script called ranges [0,1] x [0,1] x [0,1].
face_locket.py then execute this script in IPython with %run face_locket.py.
Approximate the 3-d integral
Change the circle to an ellipsoid. ∫︁ ∫︁ ∫︁
1 1 1
(𝑎𝑏 − 𝑐)𝑑𝑎 𝑑𝑏 𝑑𝑐
3.5.3 Data statistics 0 0 0

The data in populations.txt describes the populations of hares and lynxes (and carrots) in northern over this volume with the mean. The exact result is: ln 2 − 21 ≈ 0.1931 . . . — what is your relative error?
Canada during 20 years: (Hints: use elementwise operations and broadcasting. You can make np.ogrid give a number of points
>>> data = np.loadtxt('data/populations.txt') in given range with np.ogrid[0:1:20j].)
>>> year, hares, lynxes, carrots = data.T # trick: columns to variables Reminder Python functions:

>>> import matplotlib.pyplot as plt def f(a, b, c):

>>> plt.axes([0.2, 0.1, 0.5, 0.8]) return some_result
<Axes: >
>>> plt.plot(year, hares, year, lynxes, year, carrots) Solution: Python source file
[<matplotlib.lines.Line2D object at ...>, ...]
>>> plt.legend(('Hare', 'Lynx', 'Carrot'), loc=(1.05, 0.5))
<matplotlib.legend.Legend object at ...>

3.5. Some exercises 92 3.5. Some exercises 93

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

3.5.5 Mandelbrot set 3.5.6 Markov chain

Write a script that computes the Mandelbrot fractal. The Mandelbrot iteration:

N_max = 50
some_threshold = 50 Markov chain transition matrix P, and probability distribution on the states p:
1. 0 <= P[i,j] <= 1: probability to go from state i to state j
c = x + 1j*y
2. Transition rule: 𝑝𝑛𝑒𝑤 = 𝑃 𝑇 𝑝𝑜𝑙𝑑
z = 0 3. all(sum(P, axis=1) == 1), p.sum() == 1: normalization
for j in range(N_max):
z = z**2 + c Write a script that works with 5 states, and:
• Constructs a random matrix, and normalizes each row so that it is a transition matrix.
Point (x, y) belongs to the Mandelbrot set if |𝑧| < some_threshold.
• Starts from a random (normalized) probability distribution p and takes 50 steps => p_50
Do this computation by:
• Computes the stationary distribution: the eigenvector of P.T with eigenvalue 1 (numerically: closest
1. Construct a grid of c = x + 1j*y values in range [-2, 1] x [-1.5, 1.5] to 1) => p_stationary
2. Do the iteration Remember to normalize the eigenvector — I didn’t. . .
3. Form the 2-d boolean mask indicating which points are in the set • Checks if p_50 and p_stationary are equal to tolerance 1e-5
4. Save the result to an image with: Toolbox: np.random, @, np.linalg.eig, reductions, abs(), argmin, comparisons, all, np.linalg.norm,
>>> import matplotlib.pyplot as plt etc.
>>> plt.imshow(mask.T, extent=[-2, 1, -1.5, 1.5]) Solution: Python source file
<matplotlib.image.AxesImage object at ...>
>>> plt.gray()
>>> plt.savefig('mandelbrot.png')
3.6 Full code examples
Solution: Python source file
3.6.1 Full code examples for the numpy chapter
1D plotting
Plot a basic 1D figure

3.5. Some exercises 94 3.6. Full code examples 95

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

import numpy as np import numpy as np

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

x = np.linspace(0, 3, 20) rng = np.random.default_rng()

y = np.linspace(0, 9, 20) image = rng.random((30, 30))
plt.plot(x, y) plt.imshow(image, cmap="hot")
plt.plot(x, y, "o") plt.colorbar()
plt.show() plt.show()

Total running time of the script: (0 minutes 0.059 seconds) Total running time of the script: (0 minutes 0.067 seconds)

2D plotting Distances exercise

Plot a basic 2D figure Plot distances in a grid

3.6. Full code examples 96 3.6. Full code examples 97

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

import numpy as np import numpy as np

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

x, y = np.arange(5), np.arange(5)[:, np.newaxis] rng = np.random.default_rng(27446968)

distance = np.sqrt(x**2 + y**2)
plt.pcolor(distance) x = np.linspace(0, 1, 20)
plt.colorbar() y = np.cos(x) + 0.3 * rng.random(20)
plt.show() p = np.poly1d(np.polyfit(x, y, 3))

Total running time of the script: (0 minutes 0.061 seconds) t = np.linspace(0, 1, 200)
plt.plot(x, y, "o", t, p(t), "-")
plt.show()

Fitting to polynomial Total running time of the script: (0 minutes 0.045 seconds)
Plot noisy data and their polynomial fit

Fitting in Chebyshev basis

Plot noisy data and their polynomial fit in a Chebyshev basis

3.6. Full code examples 98 3.6. Full code examples 99

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

import numpy as np import numpy as np

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

rng = np.random.default_rng(27446968) data = np.loadtxt("../../../data/populations.txt")

year, hares, lynxes, carrots = data.T
x = np.linspace(-1, 1, 2000)
y = np.cos(x) + 0.3 * rng.random(2000) plt.axes((0.2, 0.1, 0.5, 0.8))
p = np.polynomial.Chebyshev.fit(x, y, 90) plt.plot(year, hares, year, lynxes, year, carrots)
plt.legend(("Hare", "Lynx", "Carrot"), loc=(1.05, 0.5))
plt.plot(x, y, "r.") plt.show()
plt.plot(x, p(x), "k-", lw=3)
plt.show() Total running time of the script: (0 minutes 0.058 seconds)

Total running time of the script: (0 minutes 0.087 seconds)

Reading and writing an elephant
Read and write images
Population exercise
Plot populations of hares, lynxes, and carrots import numpy as np
import matplotlib.pyplot as plt

original figure

plt.figure()
img = plt.imread("../../../data/elephant.png")
plt.imshow(img)

3.6. Full code examples 100 3.6. Full code examples 101
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

<matplotlib.image.AxesImage object at 0x7f24a23de3f0> <matplotlib.image.AxesImage object at 0x7f24a23c5220>

red channel displayed in grey lower resolution

plt.figure() plt.figure()
img_red = img[:, :, 0] img_tiny = img[::6, ::6]
plt.imshow(img_red, cmap="gray") plt.imshow(img_tiny, interpolation="nearest")

plt.show()

3.6. Full code examples 102 3.6. Full code examples 103
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Total running time of the script: (0 minutes 0.245 seconds) import numpy as np
import matplotlib.pyplot as plt
from numpy import newaxis
import warnings
Mandelbrot set
Compute the Mandelbrot fractal and plot it
def compute_mandelbrot(N_max, some_threshold, nx, ny):
# A grid of c-values
x = np.linspace(-2, 1, nx)
y = np.linspace(-1.5, 1.5, ny)

c = x[:, newaxis] + 1j * y[newaxis, :]

# Mandelbrot iteration

z = c

# The code below overflows in many regions of the x-y grid, suppress
# warnings temporarily
with warnings.catch_warnings():
warnings.simplefilter("ignore")
for j in range(N_max):
z = z**2 + c
mandelbrot_set = abs(z) < some_threshold

return mandelbrot_set

(continues on next page)

3.6. Full code examples 104 3.6. Full code examples 105
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

mandelbrot_set = compute_mandelbrot(50, 50.0, 601, 401) plt.figure(figsize=(4, 3))
plt.plot(t, np.sqrt(mean_sq_distance), "g.", t, np.sqrt(t), "y-")
plt.imshow(mandelbrot_set.T, extent=(-2, 1, -1.5, 1.5)) plt.xlabel(r"$t$")
plt.gray() plt.ylabel(r"$\sqrt{\langle (\delta x)^2 \rangle}$")
plt.show() plt.tight_layout()
plt.show()
Total running time of the script: (0 minutes 0.076 seconds)
Total running time of the script: (0 minutes 0.071 seconds)

Random walk exercise

Plot distance as a function of time for a random walk together with the theoretical result

import numpy as np
import matplotlib.pyplot as plt

# We create 1000 realizations with 200 steps each

n_stories = 1000
t_max = 200

t = np.arange(t_max)
# Steps can be -1 or 1 (note that randint excludes the upper limit)
rng = np.random.default_rng()
steps = 2 * rng.integers(0, 1 + 1, (n_stories, t_max)) - 1

# The time evolution of the position is obtained by successively

# summing up individual steps. This is done for each of the
# realizations, i.e. along axis 1.
positions = np.cumsum(steps, axis=1)

# Determine the time evolution of the mean square distance.

sq_distance = positions**2
mean_sq_distance = np.mean(sq_distance, axis=0)

# Plot the distance d from the origin as a function of time and

# compare with the theoretically expected result where d(t)
# grows as a square root of time t.
(continues on next page)

3.6. Full code examples 106 3.6. Full code examples 107
 Scientific Python Lectures, Edition 2025.1rc0.dev0

to visualize data from Python and publication-quality figures in many formats. We are going to
explore matplotlib in interactive mode covering most common cases.

4.1.1 IPython, Jupyter, and matplotlib modes

4
 Tip

The Jupyter notebook and the IPython enhanced interactive Python, are tuned for the scientific-
computing workflow in Python, in combination with Matplotlib:
CHAPTER
For interactive matplotlib sessions, turn on the matplotlib mode
IPython console
When using the IPython console, use:

In [1]: %matplotlib

Jupyter notebook

Matplotlib: plotting
In the notebook, insert, at the beginning of the notebook the following magic:

%matplotlib inline

4.1.2 pyplot

 Tip

pyplot provides a procedural interface to the matplotlib object-oriented plotting library. It is modeled
Thanks closely after Matlab™. Therefore, the majority of plotting commands in pyplot have Matlab™ analogs
with similar arguments. Important commands are explained with interactive examples.
Many thanks to Bill Wing and Christoph Deil for review and corrections.

import matplotlib.pyplot as plt

Authors: Nicolas Rougier, Mike Müller, Gaël Varoquaux

Chapter contents 4.2 Simple plot

• Introduction
• Simple plot  Tip

• Figures, Subplots, Axes and Ticks In this section, we want to draw the cosine and sine functions on the same plot. Starting from the
default settings, we’ll enrich the figure step by step to make it nicer.
• Other Types of Plots: examples and exercises
First step is to get the data for the sine and cosine functions:
• Beyond this tutorial
• Quick references
import numpy as np
• Full code examples
X = np.linspace(-np.pi, np.pi, 256)
C, S = np.cos(X), np.sin(X)

4.1 Introduction X is now a numpy array with 256 values ranging from −𝜋 to +𝜋 (included). C is the cosine (256 values)
and S is the sine (256 values).
To run the example, you can type them in an IPython interactive session:
 Tip
$ ipython --matplotlib
Matplotlib is probably the most used Python package for 2D-graphics. It provides both a quick way

108 4.2. Simple plot 109

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

This brings us to the IPython prompt: (continued from previous page)

IPython 0.13 -- An enhanced Interactive Python.

plt.show()
? -> Introduction to IPython's features.
%magic -> Information about IPython's 'magic' % functions.
help -> Python's own help system. 4.2.2 Instantiating defaults
object? -> Details about 'object'. ?object also works, ?? prints more.

 Tip

You can also download each of the examples and run it using regular python, but you will lose
interactive data manipulation:
$ python plot_exercise_1.py

You can get source for each step by clicking on the corresponding figure.

4.2.1 Plotting with default settings

 Hint

Documentation
• Customizing matplotlib

In the script below, we’ve instantiated (and commented) all the figure settings that influence the ap-
pearance of the plot.

 Tip
 Hint
The settings have been explicitly set to their default values, but now you can interactively play with
Documentation the values to explore their affect (see Line properties and Line styles below).
• plot tutorial
• plot() command
import numpy as np
import matplotlib.pyplot as plt
 Tip
# Create a figure of size 8x6 inches, 80 dots per inch
Matplotlib comes with a set of default settings that allow customizing all kinds of properties. You plt.figure(figsize=(8, 6), dpi=80)
can control the defaults of almost every property in matplotlib: figure size and dpi, line width, color
and style, axes, axis and grid properties, text and font properties and so on. # Create a new subplot from a grid of 1x1
plt.subplot(1, 1, 1)

X = np.linspace(-np.pi, np.pi, 256)

import numpy as np C, S = np.cos(X), np.sin(X)
import matplotlib.pyplot as plt
# Plot cosine with a blue continuous line of width 1 (pixels)
X = np.linspace(-np.pi, np.pi, 256) plt.plot(X, C, color="blue", linewidth=1.0, linestyle="-")
C, S = np.cos(X), np.sin(X)
# Plot sine with a green continuous line of width 1 (pixels)
plt.plot(X, C) plt.plot(X, S, color="green", linewidth=1.0, linestyle="-")
plt.plot(X, S)
(continues on next page)
# Set x limits
(continues on next page)

4.2. Simple plot 110 4.2. Simple plot 111

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) 4.2.4 Setting limits

plt.xlim(-4.0, 4.0)

# Set x ticks
plt.xticks(np.linspace(-4, 4, 9))

# Set y limits
plt.ylim(-1.0, 1.0)

# Set y ticks
plt.yticks(np.linspace(-1, 1, 5))

# Save figure using 72 dots per inch

# plt.savefig("exercise_2.png", dpi=72)
 Hint
# Show result on screen
plt.show() Documentation
• xlim() command

4.2.3 Changing colors and line widths • ylim() command

 Tip

Current limits of the figure are a bit too tight and we want to make some space in order to clearly
see all data points.

...
plt.xlim(X.min() * 1.1, X.max() * 1.1)
plt.ylim(C.min() * 1.1, C.max() * 1.1)
...
 Hint

Documentation
4.2.5 Setting ticks
• Controlling line properties
• Line2D API

 Tip

First step, we want to have the cosine in blue and the sine in red and a slightly thicker line for both
of them. We’ll also slightly alter the figure size to make it more horizontal.

...
plt.figure(figsize=(10, 6), dpi=80)  Hint
plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-") Documentation
... • xticks() command
• yticks() command
• Tick container

4.2. Simple plot 112 4.2. Simple plot 113

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

• Tick locating and formatting
[r'$-1$', r'$0$', r'$+1$'])
...

 Tip
4.2.7 Moving spines
Current ticks are not ideal because they do not show the interesting values (±𝜋,:math:pm pi/2) for
sine and cosine. We’ll change them such that they show only these values.

...
plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi])
plt.yticks([-1, 0, +1])
...

4.2.6 Setting tick labels

 Hint

Documentation
• spines API
• Axis container
• Transformations tutorial

 Tip

Spines are the lines connecting the axis tick marks and noting the boundaries of the data area. They
 Hint can be placed at arbitrary positions and until now, they were on the border of the axis. We’ll change
Documentation that since we want to have them in the middle. Since there are four of them (top/bottom/left/right),
we’ll discard the top and right by setting their color to none and we’ll move the bottom and left ones
• Working with text to coordinate 0 in data space coordinates.
• xticks() command
• yticks() command
• set_xticklabels() ...
ax = plt.gca() # gca stands for 'get current axis'
• set_yticklabels()
ax.spines['right'].set_color('none')
ax.spines['top'].set_color('none')
ax.xaxis.set_ticks_position('bottom')
 Tip ax.spines['bottom'].set_position(('data',0))
ax.yaxis.set_ticks_position('left')
Ticks are now properly placed but their label is not very explicit. We could guess that 3.142 is 𝜋 but ax.spines['left'].set_position(('data',0))
it would be better to make it explicit. When we set tick values, we can also provide a corresponding ...
label in the second argument list. Note that we’ll use latex to allow for nice rendering of the label.

...
plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi],
[r'$-\pi$', r'$-\pi/2$', r'$0$', r'$+\pi/2$', r'$+\pi$'])

plt.yticks([-1, 0, +1],
(continues on next page)

4.2. Simple plot 114 4.2. Simple plot 115

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

4.2.8 Adding a legend

• annotate() command

 Tip

Let’s annotate some interesting points using the annotate command. We chose the 2𝜋/3 value and
we want to annotate both the sine and the cosine. We’ll first draw a marker on the curve as well as
a straight dotted line. Then, we’ll use the annotate command to display some text with an arrow.

...

 Hint t = 2 * np.pi / 3
plt.plot([t, t], [0, np.cos(t)], color='blue', linewidth=2.5, linestyle="--")
Documentation
plt.scatter([t, ], [np.cos(t), ], 50, color='blue')
• Legend guide
plt.annotate(r'$cos(\frac{2\pi}{3} )=-\frac{1} {2} $',
• legend() command
xy=(t, np.cos(t)), xycoords='data',
• legend API xytext=(-90, -50), textcoords='offset points', fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2"))

plt.plot([t, t],[0, np.sin(t)], color='red', linewidth=2.5, linestyle="--")

 Tip plt.scatter([t, ],[np.sin(t), ], 50, color='red')
Let’s add a legend in the upper left corner. This only requires adding the keyword argument label
plt.annotate(r'$sin(\frac{2\pi}{3} )=\frac{\sqrt{3} }{2} $',
(that will be used in the legend box) to the plot commands.
xy=(t, np.sin(t)), xycoords='data',
xytext=(+10, +30), textcoords='offset points', fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2"))
...
...
plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-", label="cosine")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine") 4.2.10 Devil is in the details

plt.legend(loc='upper left')
...

4.2.9 Annotate some points

 Hint

Documentation
• artist API
• set_bbox() method
 Hint

Documentation
 Tip
• Annotating axis

4.2. Simple plot 116 4.2. Simple plot 117

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

The tick labels are now hardly visible because of the blue and red lines. We can make them bigger the current figure (no argument), (2) a specific figure (figure number or figure instance as argument),
and we can also adjust their properties such that they’ll be rendered on a semi-transparent white or (3) all figures ("all" as argument).
background. This will allow us to see both the data and the labels.

plt.close(1) # Closes figure 1

...
for label in ax.get_xticklabels() + ax.get_yticklabels():
4.3.2 Subplots
label.set_fontsize(16)
label.set_bbox(dict(facecolor='white', edgecolor='None', alpha=0.65))  Tip
...
With subplot you can arrange plots in a regular grid. You need to specify the number of rows and
columns and the number of the plot. Note that the gridspec command is a more powerful alternative.
4.3 Figures, Subplots, Axes and Ticks
A “figure” in matplotlib means the whole window in the user interface. Within this figure there can be
“subplots”.

 Tip

So far we have used implicit figure and axes creation. This is handy for fast plots. We can have more
control over the display using figure, subplot, and axes explicitly. While subplot positions the plots
in a regular grid, axes allows free placement within the figure. Both can be useful depending on your
intention. We’ve already worked with figures and subplots without explicitly calling them. When we
call plot, matplotlib calls gca() to get the current axes and gca in turn calls gcf() to get the current
figure. If there is none it calls figure() to make one, strictly speaking, to make a subplot(111).
Let’s look at the details.

4.3.1 Figures

 Tip

A figure is the windows in the GUI that has “Figure #” as title. Figures are numbered starting from
1 as opposed to the normal Python way starting from 0. This is clearly MATLAB-style. There are
several parameters that determine what the figure looks like:

Argument Default Description

num 1 number of figure
figsize figure.figsize figure size in inches (width, height)
dpi figure.dpi resolution in dots per inch 4.3.3 Axes
facecolor figure.facecolor color of the drawing background
edgecolor figure.edgecolor color of edge around the drawing background Axes are very similar to subplots but allow placement of plots at any location in the figure. So if we
frameon True draw figure frame or not want to put a smaller plot inside a bigger one we do so with axes.

 Tip

The defaults can be specified in the resource file and will be used most of the time. Only the number
of the figure is frequently changed.
As with other objects, you can set figure properties also setp or with the set_something methods.
When you work with the GUI you can close a figure by clicking on the x in the upper right corner.
But you can close a figure programmatically by calling close. Depending on the argument it closes (1)

4.3. Figures, Subplots, Axes and Ticks 118 4.3. Figures, Subplots, Axes and Ticks 119
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

4.4 Other Types of Plots: examples and exercises

4.3.4 Ticks
Well formatted ticks are an important part of publishing-ready figures. Matplotlib provides a totally
configurable system for ticks. There are tick locators to specify where ticks should appear and tick
formatters to give ticks the appearance you want. Major and minor ticks can be located and formatted
independently from each other. Per default minor ticks are not shown, i.e. there is only an empty list
for them because it is as NullLocator (see below).

Tick Locators
Tick locators control the positions of the ticks. They are set as follows:

ax = plt.gca()
ax.xaxis.set_major_locator(eval(locator))

There are several locators for different kind of requirements:

All of these locators derive from the base class matplotlib.ticker.Locator. You can make your own
locator deriving from it. Handling dates as ticks can be especially tricky. Therefore, matplotlib provides
special locators in matplotlib.dates.

4.3. Figures, Subplots, Axes and Ticks 120 4.4. Other Types of Plots: examples and exercises 121
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

4.4. Other Types of Plots: examples and exercises 122 4.4. Other Types of Plots: examples and exercises 123
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

4.4.1 Regular Plots 4.4.3 Bar Plots

Starting from the code below, try to reproduce the graphic taking care of filled areas: Starting from the code below, try to reproduce the graphic by adding labels for red bars.

 Hint  Hint

You need to use the fill_between() command. You need to take care of text alignment.

n = 256
X = np.linspace(-np.pi, np.pi, n)
n = 12
Y = np.sin(2 * X)
X = np.arange(n)
rng = np.random.default_rng()
plt.plot(X, Y + 1, color='blue', alpha=1.00)
Y1 = (1 - X / float(n)) * rng.uniform(0.5, 1.0, n)
plt.plot(X, Y - 1, color='blue', alpha=1.00)
Y2 = (1 - X / float(n)) * rng.uniform(0.5, 1.0, n)
Click on the figure for solution.
plt.bar(X, +Y1, facecolor='#9999ff', edgecolor='white')
plt.bar(X, -Y2, facecolor='#ff9999', edgecolor='white')
4.4.2 Scatter Plots
for x, y in zip(X, Y1):
plt.text(x + 0.4, y + 0.05, '%.2f ' % y, ha='center', va='bottom')

plt.ylim(-1.25, +1.25)

Click on figure for solution.

4.4.4 Contour Plots

Starting from the code below, try to reproduce the graphic taking care of marker size, color and trans-
parency.

 Hint

Color is given by angle of (X,Y).

n = 1024 Starting from the code below, try to reproduce the graphic taking care of the colormap (see Colormaps
rng = np.random.default_rng() below).
X = rng.normal(0,1,n)
Y = rng.normal(0,1,n)
 Hint
plt.scatter(X,Y)
You need to use the clabel() command.
Click on figure for solution.

4.4. Other Types of Plots: examples and exercises 124 4.4. Other Types of Plots: examples and exercises 125
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Starting from the code below, try to reproduce the graphic taking care of colors and slices size.
def f(x, y):
return (1 - x / 2 + x ** 5 + y ** 3) * np.exp(-x ** 2 -y ** 2)
 Hint
n = 256
x = np.linspace(-3, 3, n) You need to modify Z.
y = np.linspace(-3, 3, n)
X, Y = np.meshgrid(x, y)
rng = np.random.default_rng()
plt.contourf(X, Y, f(X, Y), 8, alpha=.75, cmap='jet') Z = rng.uniform(0, 1, 20)
C = plt.contour(X, Y, f(X, Y), 8, colors='black', linewidth=.5) plt.pie(Z)

Click on figure for solution. Click on the figure for the solution.

4.4.5 Imshow 4.4.7 Quiver Plots

Starting from the code below, try to reproduce the graphic taking care of colormap, image interpolation Starting from the code below, try to reproduce the graphic taking care of colors and orientations.
and origin.
 Hint
 Hint
You need to draw arrows twice.
You need to take care of the origin of the image in the imshow command and use a colorbar()
n = 8
X, Y = np.mgrid[0:n, 0:n]
def f(x, y):
plt.quiver(X, Y)
return (1 - x / 2 + x ** 5 + y ** 3) * np.exp(-x ** 2 - y ** 2)

n = 10 Click on figure for solution.

x = np.linspace(-3, 3, 4 * n)
y = np.linspace(-3, 3, 3 * n) 4.4.8 Grids
X, Y = np.meshgrid(x, y)
plt.imshow(f(X, Y))

Click on the figure for the solution.

4.4.6 Pie Charts

Starting from the code below, try to reproduce the graphic taking care of line styles.

axes = plt.gca()
axes.set_xlim(0, 4)
axes.set_ylim(0, 3)
(continues on next page)

4.4. Other Types of Plots: examples and exercises 126 4.4. Other Types of Plots: examples and exercises 127
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

axes.set_xticklabels([]) radii = 10 * rng.random(N)
axes.set_yticklabels([]) width = np.pi / 4 * rng.random(N)
bars = plt.bar(theta, radii, width=width, bottom=0.0)
Click on figure for solution.
for r, bar in zip(radii, bars):
4.4.9 Multi Plots bar.set_facecolor(plt.cm.jet(r / 10.))
bar.set_alpha(0.5)

Click on figure for solution.

4.4.11 3D Plots

Starting from the code below, try to reproduce the graphic.

 Hint

You can use several subplots with different partition.

Starting from the code below, try to reproduce the graphic.

plt.subplot(2, 2, 1)
 Hint
plt.subplot(2, 2, 3)
plt.subplot(2, 2, 4) You need to use contourf()

Click on figure for solution.

from mpl_toolkits.mplot3d import Axes3D
4.4.10 Polar Axis
fig = plt.figure()
ax = Axes3D(fig)
X = np.arange(-4, 4, 0.25)
Y = np.arange(-4, 4, 0.25)
X, Y = np.meshgrid(X, Y)
R = np.sqrt(X**2 + Y**2)
Z = np.sin(R)

ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap='hot')

Click on figure for solution.

4.4.12 Text
 Hint

You only need to modify the axes line

Starting from the code below, try to reproduce the graphic.

plt.axes([0, 0, 1, 1])

N = 20
theta = np.arange(0., 2 * np.pi, 2 * np.pi / N)
rng = np.random.default_rng()
(continues on next page)

4.4. Other Types of Plots: examples and exercises 128 4.4. Other Types of Plots: examples and exercises 129
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Try to do the same from scratch ! 4.5.2 Matplotlib documentation

• User guide • How-To

 Hint • FAQ • Troubleshooting
Have a look at the matplotlib logo. • Installation • Environment Variables
• Usage • Screenshots

Click on figure for solution.

4.5.3 Code documentation
The code is well documented and you can quickly access a specific command from within a python
session:

>>> import matplotlib.pyplot as plt

>>> help(plt.plot)
Help on function plot in module matplotlib.pyplot:

plot(*args: ...) -> 'list[Line2D]'

Plot y versus x as lines and/or markers.

Quick read Call signatures::

If you want to do a first quick pass through the Scientific Python Lectures to learn the ecosystem, plot([x], y, [fmt], *, data=None, **kwargs)
you can directly skip to the next chapter: SciPy : high-level scientific computing. plot([x], y, [fmt], [x2], y2, [fmt2], ..., **kwargs)
...
The remainder of this chapter is not necessary to follow the rest of the intro part. But be sure to
come back and finish this chapter later.
4.5.4 Galleries
The matplotlib gallery is also incredibly useful when you search how to render a given graphic. Each
4.5 Beyond this tutorial example comes with its source.
Matplotlib benefits from extensive documentation as well as a large community of users and developers.
Here are some links of interest: 4.5.5 Mailing lists
Finally, there is a user mailing list where you can ask for help and a developers mailing list that is more
4.5.1 Tutorials technical.

• Pyplot tutorial • Customizing your objects

• Introduction • Object containers 4.6 Quick references
• Controlling line properties • Figure container
• Working with multiple figures and axes • Axes container Here is a set of tables that show main properties and styles.
• Working with text • Axis containers
• Image tutorial • Tick containers
• Startup commands • Path tutorial
• Importing image data into NumPy arrays • Introduction
• Plotting NumPy arrays as images • Bézier example
• Text tutorial • Compound paths
• Text introduction • Transforms tutorial
• Basic text commands • Introduction
• Text properties and layout • Data coordinates
• Writing mathematical expressions • Axes coordinates
• Text rendering With LaTeX • Blended transformations
• Annotating text • Using offset transforms to create a shadow
• Artist tutorial effect
• Introduction • The transformation pipeline

4.5. Beyond this tutorial 130 4.6. Quick references 131

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

4.6.1 Line properties 4.6.3 Markers

Property Description Appearance

alpha (or a) alpha transparency on 0-1

scale
antialiased True or False - use antialised
rendering

color (or c) matplotlib color arg

linestyle (or ls) see Line properties
linewidth (or lw) float, the line width in points
solid_capstyle Cap style for solid lines
solid_joinstyle Join style for solid lines
dash_capstyle Cap style for dashes
dash_joinstyle Join style for dashes
marker see Markers
markeredgewidth line width around the
(mew) marker symbol 4.6.4 Colormaps
markeredgecolor edge color if a marker is used All colormaps can be reversed by appending _r. For instance, gray_r is the reverse of gray.
(mec) If you want to know more about colormaps, check the documentation on Colormaps in matplotlib.
markerfacecolor face color if a marker is used
(mfc)
markersize (ms) size of the marker in points

4.6.2 Line styles

4.7 Full code examples

4.7.1 Code samples for Matplotlib
The examples here are only examples relevant to the points raised in this chapter. The matplotlib
documentation comes with a much more exhaustive gallery.

4.6. Quick references 132 4.7. Full code examples 133

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Pie chart
A simple pie chart example with matplotlib.

import numpy as np
import matplotlib

matplotlib.use("Agg")
import matplotlib.pyplot as plt

fig = plt.figure(figsize=(5, 4), dpi=72)

import numpy as np axes = fig.add_axes((0.01, 0.01, 0.98, 0.98))
import matplotlib.pyplot as plt X = np.linspace(0, 2, 200)
Y = np.sin(2 * np.pi * X)
n = 20 plt.plot(X, Y, lw=2)
Z = np.ones(n) plt.ylim(-1.1, 1.1)
Z[-1] *= 2 plt.grid()

plt.axes((0.025, 0.025, 0.95, 0.95)) plt.show()

plt.pie(Z, explode=Z * 0.05, colors=[f"{ i / float(n): f} " for i in range(n)]) Total running time of the script: (0 minutes 0.052 seconds)
plt.axis("equal")
plt.xticks([])
plt.yticks() Plotting a scatter of points
plt.show() A simple example showing how to plot a scatter of points with matplotlib.

Total running time of the script: (0 minutes 0.051 seconds)

A simple, good-looking plot

Demoing some simple features of matplotlib

4.7. Full code examples 134 4.7. Full code examples 135
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Subplots
Show multiple subplots in matplotlib.

import numpy as np
import matplotlib.pyplot as plt

n = 1024 import matplotlib.pyplot as plt

rng = np.random.default_rng()
X = rng.normal(0, 1, n) fig = plt.figure()
Y = rng.normal(0, 1, n) fig.subplots_adjust(bottom=0.025, left=0.025, top=0.975, right=0.975)
T = np.arctan2(Y, X)
plt.subplot(2, 1, 1)
plt.axes((0.025, 0.025, 0.95, 0.95)) plt.xticks([]), plt.yticks([])
plt.scatter(X, Y, s=75, c=T, alpha=0.5)
plt.subplot(2, 3, 4)
plt.xlim(-1.5, 1.5) plt.xticks([])
plt.xticks([]) plt.yticks([])
plt.ylim(-1.5, 1.5)
plt.yticks([]) plt.subplot(2, 3, 5)
plt.xticks([])
plt.show() plt.yticks([])

Total running time of the script: (0 minutes 0.058 seconds) plt.subplot(2, 3, 6)

plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.039 seconds)

4.7. Full code examples 136 4.7. Full code examples 137
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Horizontal arrangement of subplots

An example showing horizontal arrangement of subplots with matplotlib.

import numpy as np
import matplotlib

import matplotlib.pyplot as plt matplotlib.use("Agg")

import matplotlib.pyplot as plt
plt.figure(figsize=(6, 4))
plt.subplot(2, 1, 1) fig = plt.figure(figsize=(5, 4), dpi=72)
plt.xticks([]) axes = fig.add_axes((0.01, 0.01, 0.98, 0.98))
plt.yticks([]) x = np.linspace(0, 2, 200)
plt.text(0.5, 0.5, "subplot(2,1,1)", ha="center", va="center", size=24, alpha=0.5) y = np.sin(2 * np.pi * x)
plt.plot(x, y, lw=0.25, c="k")
plt.subplot(2, 1, 2) plt.xticks(np.arange(0.0, 2.0, 0.1))
plt.xticks([]) plt.yticks(np.arange(-1.0, 1.0, 0.1))
plt.yticks([]) plt.grid()
plt.text(0.5, 0.5, "subplot(2,1,2)", ha="center", va="center", size=24, alpha=0.5) plt.show()

plt.tight_layout() Total running time of the script: (0 minutes 0.085 seconds)

plt.show()

Total running time of the script: (0 minutes 0.042 seconds)

Subplot plot arrangement vertical
An example showing vertical arrangement of subplots with matplotlib.
A simple plotting example
A plotting example with a few simple tweaks

4.7. Full code examples 138 4.7. Full code examples 139
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

import matplotlib.pyplot as plt

plt.figure(figsize=(6, 4))
plt.subplot(1, 2, 1) import matplotlib.pyplot as plt
plt.xticks([])
plt.yticks([]) plt.axes((0.1, 0.1, 0.8, 0.8))
plt.text(0.5, 0.5, "subplot(1,2,1)", ha="center", va="center", size=24, alpha=0.5) plt.xticks([])
plt.yticks([])
plt.subplot(1, 2, 2) plt.text(
plt.xticks([]) 0.6, 0.6, "axes([0.1, 0.1, 0.8, 0.8])", ha="center", va="center", size=20,␣
plt.yticks([]) ˓→alpha=0.5
plt.text(0.5, 0.5, "subplot(1,2,2)", ha="center", va="center", size=24, alpha=0.5) )

plt.tight_layout() plt.axes((0.2, 0.2, 0.3, 0.3))

plt.show() plt.xticks([])
plt.yticks([])
Total running time of the script: (0 minutes 0.034 seconds) plt.text(
0.5, 0.5, "axes([0.2, 0.2, 0.3, 0.3])", ha="center", va="center", size=16,␣
˓→alpha=0.5

)
Simple axes example
This example shows a couple of simple usage of axes. plt.show()

Total running time of the script: (0 minutes 0.042 seconds)

3D plotting
A simple example of 3D plotting.

4.7. Full code examples 140 4.7. Full code examples 141
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

import numpy as np import numpy as np

import matplotlib.pyplot as plt import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D

ax: Axes3D = plt.figure().add_subplot(projection="3d") def f(x, y):

x = np.arange(-4, 4, 0.25) return (1 - x / 2 + x**5 + y**3) * np.exp(-(x**2) - y**2)
y = np.arange(-4, 4, 0.25)
X, Y = np.meshgrid(x, y)
R = np.sqrt(X**2 + Y**2) n = 10
Z = np.sin(R) x = np.linspace(-3, 3, int(3.5 * n))
y = np.linspace(-3, 3, int(3.0 * n))
ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap="hot") X, Y = np.meshgrid(x, y)
ax.contourf(X, Y, Z, zdir="z", offset=-2, cmap="hot") Z = f(X, Y)
ax.set_zlim(-2, 2)
plt.axes((0.025, 0.025, 0.95, 0.95))
plt.show() plt.imshow(Z, interpolation="nearest", cmap="bone", origin="lower")
plt.colorbar(shrink=0.92)
Total running time of the script: (0 minutes 0.140 seconds)
plt.xticks([])
plt.yticks([])
plt.show()
Imshow elaborate
An example demoing imshow and styling the figure. Total running time of the script: (0 minutes 0.058 seconds)

4.7. Full code examples 142 4.7. Full code examples 143
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Plotting a vector field: quiver Displaying the contours of a function

A simple example showing how to plot a vector field (quiver) with matplotlib. An example showing how to display the contours of a function with matplotlib.

import numpy as np import numpy as np

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

n = 8
X, Y = np.mgrid[0:n, 0:n] def f(x, y):
T = np.arctan2(Y - n / 2.0, X - n / 2.0) return (1 - x / 2 + x**5 + y**3) * np.exp(-(x**2) - y**2)
R = 10 + np.sqrt((Y - n / 2.0) ** 2 + (X - n / 2.0) ** 2)
U, V = R * np.cos(T), R * np.sin(T)
n = 256
plt.axes((0.025, 0.025, 0.95, 0.95)) x = np.linspace(-3, 3, n)
plt.quiver(X, Y, U, V, R, alpha=0.5) y = np.linspace(-3, 3, n)
plt.quiver(X, Y, U, V, edgecolor="k", facecolor="None", linewidth=0.5) X, Y = np.meshgrid(x, y)

plt.xlim(-1, n) plt.axes((0.025, 0.025, 0.95, 0.95))

plt.xticks([])
plt.ylim(-1, n) plt.contourf(X, Y, f(X, Y), 8, alpha=0.75, cmap="hot")
plt.yticks([]) C = plt.contour(X, Y, f(X, Y), 8, colors="black", linewidths=0.5)
plt.clabel(C, inline=1, fontsize=10)
plt.show()
plt.xticks([])
Total running time of the script: (0 minutes 0.029 seconds) plt.yticks([])
plt.show()

Total running time of the script: (0 minutes 0.089 seconds)

4.7. Full code examples 144 4.7. Full code examples 145
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

A example of plotting not quite right

An “ugly” example of plotting.

import numpy as np
import matplotlib import numpy as np
import matplotlib.pyplot as plt
matplotlib.use("Agg")
import matplotlib.pyplot as plt n = 256
X = np.linspace(-np.pi, np.pi, n)
matplotlib.rc("grid", color="black", linestyle="-", linewidth=1) Y = np.sin(2 * X)

fig = plt.figure(figsize=(5, 4), dpi=72) plt.axes((0.025, 0.025, 0.95, 0.95))

axes = fig.add_axes((0.01, 0.01, 0.98, 0.98), facecolor=".75")
X = np.linspace(0, 2, 40) plt.plot(X, Y + 1, color="blue", alpha=1.00)
Y = np.sin(2 * np.pi * X) plt.fill_between(X, 1, Y + 1, color="blue", alpha=0.25)
plt.plot(X, Y, lw=0.05, c="b", antialiased=False)
plt.plot(X, Y - 1, color="blue", alpha=1.00)
plt.xticks([]) plt.fill_between(X, -1, Y - 1, (Y - 1) > -1, color="blue", alpha=0.25)
plt.yticks(np.arange(-1.0, 1.0, 0.2)) plt.fill_between(X, -1, Y - 1, (Y - 1) < -1, color="red", alpha=0.25)
plt.grid()
ax = plt.gca() plt.xlim(-np.pi, np.pi)
plt.xticks([])
plt.show() plt.ylim(-2.5, 2.5)
plt.yticks([])
Total running time of the script: (0 minutes 0.032 seconds)
plt.show()

Total running time of the script: (0 minutes 0.031 seconds)

Plot and filled plots
Simple example of plots and filling between them with matplotlib.

4.7. Full code examples 146 4.7. Full code examples 147
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Plotting in polar coordinates

A simple example showing how to plot in polar coordinates with matplotlib.
Bar plots
An example of bar plots with matplotlib.

import numpy as np

import matplotlib import numpy as np

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

n = 12
jet = matplotlib.colormaps["jet"] X = np.arange(n)
rng = np.random.default_rng()
ax = plt.axes((0.025, 0.025, 0.95, 0.95), polar=True) Y1 = (1 - X / n) * rng.uniform(0.5, 1.0, n)
Y2 = (1 - X / n) * rng.uniform(0.5, 1.0, n)
N = 20
theta = np.arange(0.0, 2 * np.pi, 2 * np.pi / N) plt.axes((0.025, 0.025, 0.95, 0.95))
rng = np.random.default_rng() plt.bar(X, +Y1, facecolor="#9999ff", edgecolor="white")
radii = 10 * rng.random(N) plt.bar(X, -Y2, facecolor="#ff9999", edgecolor="white")
width = np.pi / 4 * rng.random(N)
bars = plt.bar(theta, radii, width=width, bottom=0.0) for x, y in zip(X, Y1):
plt.text(x, y + 0.05, f"{ y: .2f} ", ha="center", va="bottom")
for r, bar in zip(radii, bars, strict=True):
bar.set_facecolor(jet(r / 10.0)) for x, y in zip(X, Y2):
bar.set_alpha(0.5) plt.text(x, -y - 0.05, f"{ y: .2f} ", ha="center", va="top")

ax.set_xticklabels([]) plt.xlim(-0.5, n)
ax.set_yticklabels([]) plt.xticks([])
plt.show() plt.ylim(-1.25, 1.25)
plt.yticks([])
Total running time of the script: (0 minutes 0.078 seconds) (continues on next page)

4.7. Full code examples 148 4.7. Full code examples 149
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

plt.yticks([])
plt.show() plt.text(0.5, 0.5, "subplot(2,2,4)", ha="center", va="center", size=20, alpha=0.5)

Total running time of the script: (0 minutes 0.050 seconds) plt.tight_layout()

plt.show()

Total running time of the script: (0 minutes 0.055 seconds)

Subplot grid
An example showing the subplot grid in matplotlib.
Axes
This example shows various axes command to position matplotlib axes.

import matplotlib.pyplot as plt

plt.figure(figsize=(6, 4))
plt.subplot(2, 2, 1)
plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "subplot(2,2,1)", ha="center", va="center", size=20, alpha=0.5)
import matplotlib.pyplot as plt
plt.subplot(2, 2, 2)
plt.xticks([]) plt.axes((0.1, 0.1, 0.5, 0.5))
plt.yticks([]) plt.xticks([])
plt.text(0.5, 0.5, "subplot(2,2,2)", ha="center", va="center", size=20, alpha=0.5) plt.yticks([])
plt.text(
plt.subplot(2, 2, 3) 0.1, 0.1, "axes((0.1, 0.1, 0.5, 0.5))", ha="left", va="center", size=16, alpha=0.5
plt.xticks([]) )
plt.yticks([])
plt.axes((0.2, 0.2, 0.5, 0.5))
plt.text(0.5, 0.5, "subplot(2,2,3)", ha="center", va="center", size=20, alpha=0.5) plt.xticks([])
plt.yticks([])
plt.subplot(2, 2, 4) plt.text(
plt.xticks([]) 0.1, 0.1, "axes((0.2, 0.2, 0.5, 0.5))", ha="left", va="center", size=16, alpha=0.5
(continues on next page) (continues on next page)

4.7. Full code examples 150 4.7. Full code examples 151
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

)
ax = plt.axes((0.025, 0.025, 0.95, 0.95))
plt.axes((0.3, 0.3, 0.5, 0.5))
plt.xticks([]) ax.set_xlim(0, 4)
plt.yticks([]) ax.set_ylim(0, 3)
plt.text( ax.xaxis.set_major_locator(ticker.MultipleLocator(1.0))
0.1, 0.1, "axes((0.3, 0.3, 0.5, 0.5))", ha="left", va="center", size=16, alpha=0.5 ax.xaxis.set_minor_locator(ticker.MultipleLocator(0.1))
) ax.yaxis.set_major_locator(ticker.MultipleLocator(1.0))
ax.yaxis.set_minor_locator(ticker.MultipleLocator(0.1))
plt.axes((0.4, 0.4, 0.5, 0.5)) ax.grid(which="major", axis="x", linewidth=0.75, linestyle="-", color="0.75")
plt.xticks([]) ax.grid(which="minor", axis="x", linewidth=0.25, linestyle="-", color="0.75")
plt.yticks([]) ax.grid(which="major", axis="y", linewidth=0.75, linestyle="-", color="0.75")
plt.text( ax.grid(which="minor", axis="y", linewidth=0.25, linestyle="-", color="0.75")
0.1, 0.1, "axes((0.4, 0.4, 0.5, 0.5))", ha="left", va="center", size=16, alpha=0.5 ax.set_xticklabels([])
) ax.set_yticklabels([])

plt.show() plt.show()

Total running time of the script: (0 minutes 0.044 seconds) Total running time of the script: (0 minutes 0.093 seconds)

Grid 3D plotting
Displaying a grid on the axes in matploblib. Demo 3D plotting with matplotlib and style the figure.

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

from matplotlib import ticker from mpl_toolkits.mplot3d.axes3d import Axes3D, get_test_data
(continues on next page) (continues on next page)

4.7. Full code examples 152 4.7. Full code examples 153
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

ax: Axes3D = plt.figure().add_subplot(projection="3d")

X, Y, Z = get_test_data(0.05)
cset = ax.contourf(X, Y, Z)
ax.clabel(cset, fontsize=9, inline=1)

plt.xticks([])
plt.yticks([])
ax.set_zticks([])

ax.text2D(
-0.05,
1.05,
" 3D plots \n",
horizontalalignment="left",
verticalalignment="top",
bbox={"facecolor": "white", "alpha": 1.0},
family="DejaVu Sans",
size="x-large",
transform=plt.gca().transAxes,
)

ax.text2D(
-0.05, import matplotlib.pyplot as plt
0.975, from matplotlib import gridspec
" Plot 2D or 3D data",
horizontalalignment="left", plt.figure(figsize=(6, 4))
verticalalignment="top", G = gridspec.GridSpec(3, 3)
family="DejaVu Sans",
size="medium", axes_1 = plt.subplot(G[0, :])
transform=plt.gca().transAxes, plt.xticks([])
) plt.yticks([])
plt.text(0.5, 0.5, "Axes 1", ha="center", va="center", size=24, alpha=0.5)
plt.show()
axes_2 = plt.subplot(G[1, :-1])
Total running time of the script: (0 minutes 0.050 seconds) plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "Axes 2", ha="center", va="center", size=24, alpha=0.5)

GridSpec axes_3 = plt.subplot(G[1:, -1])

plt.xticks([])
An example demoing gridspec
plt.yticks([])
plt.text(0.5, 0.5, "Axes 3", ha="center", va="center", size=24, alpha=0.5)

axes_4 = plt.subplot(G[-1, 0])

plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "Axes 4", ha="center", va="center", size=24, alpha=0.5)

axes_5 = plt.subplot(G[-1, -2])

plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "Axes 5", ha="center", va="center", size=24, alpha=0.5)

plt.tight_layout()
plt.show()

4.7. Full code examples 154 4.7. Full code examples 155
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Total running time of the script: (0 minutes 0.069 seconds) (continued from previous page)
rng = np.random.default_rng()

for i in range(24):
Demo text printing
index = rng.integers(0, len(eqs))
A example showing off elaborate text printing with matplotlib. eq = eqs[index]
size = np.random.uniform(12, 32)
x, y = np.random.uniform(0, 1, 2)
alpha = np.random.uniform(0.25, 0.75)
plt.text(
x,
y,
eq,
ha="center",
va="center",
color="#11557c",
alpha=alpha,
transform=plt.gca().transAxes,
fontsize=size,
clip_on=True,
)
plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.651 seconds)

4.7.2 Code for the chapter’s exercises

4.7.3 Example demoing choices for an option
4.7.4 Code generating the summary figures with a title
Code for the chapter’s exercises
import numpy as np
import matplotlib.pyplot as plt Exercise 1

Solution of the exercise 1 with matplotlib.

eqs = []
eqs.append(
r"$W^{3\beta}_{\delta_1 \rho_1 \sigma_2} = U^{3\beta}_{\delta_1 \rho_1} + \frac{1}
˓→ {8 \pi 2} \int^{\alpha_2}_{\alpha_2} d \alpha^\prime_2 \left[\frac{ U^{2\beta}_{\

˓→delta_1 \rho_1} - \alpha^\prime_2U^{1\beta}_{\rho_1 \sigma_2} }{U^{0\beta}_{\rho_1 \

˓→sigma_2}}\right]$"

)
eqs.append(
r"$\frac{d\rho}{d t} + \rho \vec{v} \cdot\nabla\vec{v} = -\nabla p + \mu\nabla^2 \
˓→vec{v} + \rho \vec{g} $"

)
eqs.append(r"$\int_{-\infty}^\infty e^{-x^2}dx=\sqrt{\pi}$")
eqs.append(r"$E = mc^2 = \sqrt{{m_0}^2c^4 + p^2c^2}$")
eqs.append(r"$F_G = G\frac{m_1m_2} {r^2}$")

plt.axes((0.025, 0.025, 0.95, 0.95))

(continues on next page)

4.7. Full code examples 156 4.7. Full code examples 157
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

import numpy as np
import matplotlib.pyplot as plt

plt.figure(figsize=(8, 5), dpi=80)

plt.subplot(111)
import numpy as np
import matplotlib.pyplot as plt X = np.linspace(-np.pi, np.pi, 256)
S = np.sin(X)
n = 256 C = np.cos(X)
X = np.linspace(-np.pi, np.pi, 256)
C, S = np.cos(X), np.sin(X) plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-")
plt.plot(X, C) plt.plot(X, S, color="red", linewidth=2.5, linestyle="-")
plt.plot(X, S)
plt.xlim(X.min() * 1.1, X.max() * 1.1)
plt.show() plt.ylim(C.min() * 1.1, C.max() * 1.1)

Total running time of the script: (0 minutes 0.056 seconds) plt.show()

Total running time of the script: (0 minutes 0.053 seconds)

Exercise 4

Exercise 4 with matplotlib.

Exercise 3

Exercise 3 with matplotlib.

4.7. Full code examples 158 4.7. Full code examples 159
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

import numpy as np import numpy as np

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

plt.figure(figsize=(8, 5), dpi=80) plt.figure(figsize=(8, 5), dpi=80)

plt.subplot(111) plt.subplot(111)

X = np.linspace(-np.pi, np.pi, 256) X = np.linspace(-np.pi, np.pi, 256)

C, S = np.cos(X), np.sin(X) S = np.sin(X)
C = np.cos(X)
plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-") plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-")
plt.xlim(-4.0, 4.0)
plt.xticks(np.linspace(-4, 4, 9)) plt.xlim(X.min() * 1.1, X.max() * 1.1)
plt.xticks([-np.pi, -np.pi / 2, 0, np.pi / 2, np.pi])
plt.ylim(-1.0, 1.0)
plt.yticks(np.linspace(-1, 1, 5)) plt.ylim(C.min() * 1.1, C.max() * 1.1)
plt.yticks([-1, 0, +1])
plt.show()
plt.show()
Total running time of the script: (0 minutes 0.046 seconds)
Total running time of the script: (0 minutes 0.037 seconds)

Exercise 5
Exercise 6
Exercise 5 with matplotlib.
Exercise 6 with matplotlib.

4.7. Full code examples 160 4.7. Full code examples 161
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Exercise 2

Exercise 2 with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

plt.figure(figsize=(8, 5), dpi=80)

plt.subplot(111)

X = np.linspace(-np.pi, np.pi, 256)

C = np.cos(X)
S = np.sin(X) import numpy as np
import matplotlib.pyplot as plt
plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-") # Create a new figure of size 8x6 points, using 100 dots per inch
plt.figure(figsize=(8, 6), dpi=80)
plt.xlim(X.min() * 1.1, X.max() * 1.1)
plt.xticks( # Create a new subplot from a grid of 1x1
[-np.pi, -np.pi / 2, 0, np.pi / 2, np.pi], plt.subplot(111)
[r"$-\pi$", r"$-\pi/2$", r"$0$", r"$+\pi/2$", r"$+\pi$"],
) X = np.linspace(-np.pi, np.pi, 256)
C, S = np.cos(X), np.sin(X)
plt.ylim(C.min() * 1.1, C.max() * 1.1)
plt.yticks([-1, 0, +1], [r"$-1$", r"$0$", r"$+1$"]) # Plot cosine using blue color with a continuous line of width 1 (pixels)
plt.plot(X, C, color="blue", linewidth=1.0, linestyle="-")
plt.show()
# Plot sine using green color with a continuous line of width 1 (pixels)
Total running time of the script: (0 minutes 0.049 seconds) plt.plot(X, S, color="green", linewidth=1.0, linestyle="-")

# Set x limits
plt.xlim(-4.0, 4.0)

# Set x ticks
plt.xticks(np.linspace(-4, 4, 9))

# Set y limits
(continues on next page)

4.7. Full code examples 162 4.7. Full code examples 163
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

plt.ylim(-1.0, 1.0) ax.xaxis.set_ticks_position("bottom")
ax.spines["bottom"].set_position(("data", 0))
# Set y ticks ax.yaxis.set_ticks_position("left")
plt.yticks(np.linspace(-1, 1, 5)) ax.spines["left"].set_position(("data", 0))

# Show result on screen plt.xlim(X.min() * 1.1, X.max() * 1.1)

plt.show() plt.xticks(
[-np.pi, -np.pi / 2, 0, np.pi / 2, np.pi],
Total running time of the script: (0 minutes 0.049 seconds) [r"$-\pi$", r"$-\pi/2$", r"$0$", r"$+\pi/2$", r"$+\pi$"],
)

plt.ylim(C.min() * 1.1, C.max() * 1.1)

Exercise 7 plt.yticks([-1, 0, +1], [r"$-1$", r"$0$", r"$+1$"])
Exercise 7 with matplotlib
plt.show()

Total running time of the script: (0 minutes 0.053 seconds)

Exercise 8

Exercise 8 with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

plt.figure(figsize=(8, 5), dpi=80)

plt.subplot(111)

X = np.linspace(-np.pi, np.pi, 256, endpoint=True)

C = np.cos(X)
S = np.sin(X) import numpy as np
import matplotlib.pyplot as plt
plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-") plt.figure(figsize=(8, 5), dpi=80)
plt.subplot(111)
ax = plt.gca()
ax.spines["right"].set_color("none") X = np.linspace(-np.pi, np.pi, 256, endpoint=True)
ax.spines["top"].set_color("none") C = np.cos(X)
(continues on next page) (continues on next page)

4.7. Full code examples 164 4.7. Full code examples 165
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

import numpy as np
S = np.sin(X) import matplotlib.pyplot as plt

plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-", label="cosine") plt.figure(figsize=(8, 5), dpi=80)

plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine") plt.subplot(111)

ax = plt.gca() X = np.linspace(-np.pi, np.pi, 256, endpoint=True)

ax.spines["right"].set_color("none") C = np.cos(X)
ax.spines["top"].set_color("none") S = np.sin(X)
ax.xaxis.set_ticks_position("bottom")
ax.spines["bottom"].set_position(("data", 0)) plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-", label="cosine")
ax.yaxis.set_ticks_position("left") plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine")
ax.spines["left"].set_position(("data", 0))
ax = plt.gca()
plt.xlim(X.min() * 1.1, X.max() * 1.1) ax.spines["right"].set_color("none")
plt.xticks( ax.spines["top"].set_color("none")
[-np.pi, -np.pi / 2, 0, np.pi / 2, np.pi], ax.xaxis.set_ticks_position("bottom")
[r"$-\pi$", r"$-\pi/2$", r"$0$", r"$+\pi/2$", r"$+\pi$"], ax.spines["bottom"].set_position(("data", 0))
) ax.yaxis.set_ticks_position("left")
ax.spines["left"].set_position(("data", 0))
plt.ylim(C.min() * 1.1, C.max() * 1.1)
plt.yticks([-1, +1], [r"$-1$", r"$+1$"]) plt.xlim(X.min() * 1.1, X.max() * 1.1)
plt.xticks(
plt.legend(loc="upper left") [-np.pi, -np.pi / 2, 0, np.pi / 2, np.pi],
[r"$-\pi$", r"$-\pi/2$", r"$0$", r"$+\pi/2$", r"$+\pi$"],
plt.show() )

Total running time of the script: (0 minutes 0.058 seconds) plt.ylim(C.min() * 1.1, C.max() * 1.1)
plt.yticks([-1, +1], [r"$-1$", r"$+1$"])

t = 2 * np.pi / 3
Exercise 9
plt.plot([t, t], [0, np.cos(t)], color="blue", linewidth=1.5, linestyle="--")
Exercise 9 with matplotlib. plt.scatter(
[
t,
],
[
np.cos(t),
],
50,
color="blue",
)
plt.annotate(
r"$sin(\frac{2\pi}{3} )=\frac{\sqrt{3} }{2} $",
xy=(t, np.sin(t)),
xycoords="data",
xytext=(+10, +30),
textcoords="offset points",
fontsize=16,
arrowprops={"arrowstyle": "->", "connectionstyle": "arc3,rad=.2"},
)

plt.plot([t, t], [0, np.sin(t)], color="red", linewidth=1.5, linestyle="--")

plt.scatter(
[
t,
],
(continues on next page)

4.7. Full code examples 166 4.7. Full code examples 167
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

[ plt.subplot(111)
np.sin(t),
], X = np.linspace(-np.pi, np.pi, 256)
50, C, S = np.cos(X), np.sin(X)
color="red",
) plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-", label="cosine")
plt.annotate( plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine")
r"$cos(\frac{2\pi}{3} )=-\frac{1} {2} $",
xy=(t, np.cos(t)), ax = plt.gca()
xycoords="data", ax.spines["right"].set_color("none")
xytext=(-90, -50), ax.spines["top"].set_color("none")
textcoords="offset points", ax.xaxis.set_ticks_position("bottom")
fontsize=16, ax.spines["bottom"].set_position(("data", 0))
arrowprops={"arrowstyle": "->", "connectionstyle": "arc3,rad=.2"}, ax.yaxis.set_ticks_position("left")
) ax.spines["left"].set_position(("data", 0))

plt.legend(loc="upper left") plt.xlim(X.min() * 1.1, X.max() * 1.1)

plt.xticks(
plt.show() [-np.pi, -np.pi / 2, 0, np.pi / 2, np.pi],
[r"$-\pi$", r"$-\pi/2$", r"$0$", r"$+\pi/2$", r"$+\pi$"],
Total running time of the script: (0 minutes 0.095 seconds) )

plt.ylim(C.min() * 1.1, C.max() * 1.1)

plt.yticks([-1, 1], [r"$-1$", r"$+1$"])
Exercise
plt.legend(loc="upper left")
Exercises with matplotlib.
t = 2 * np.pi / 3
plt.plot([t, t], [0, np.cos(t)], color="blue", linewidth=1.5, linestyle="--")
plt.scatter(
[
t,
],
[
np.cos(t),
],
50,
color="blue",
)
plt.annotate(
r"$sin(\frac{2\pi}{3} )=\frac{\sqrt{3} }{2} $",
xy=(t, np.sin(t)),
xycoords="data",
xytext=(10, 30),
textcoords="offset points",
fontsize=16,
arrowprops={"arrowstyle": "->", "connectionstyle": "arc3,rad=.2"},
)

plt.plot([t, t], [0, np.sin(t)], color="red", linewidth=1.5, linestyle="--")

plt.scatter(
[
import numpy as np t,
import matplotlib.pyplot as plt ],
[
plt.figure(figsize=(8, 5), dpi=80) np.sin(t),
(continues on next page)
(continues on next page)

4.7. Full code examples 168 4.7. Full code examples 169
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) Linewidth

],
50, Plot various linewidth with matplotlib.
color="red",
)
plt.annotate(
import matplotlib.pyplot as plt
r"$cos(\frac{2\pi}{3} )=-\frac{1} {2} $",
xy=(t, np.cos(t)),
size = 256, 16
xycoords="data",
dpi = 72.0
xytext=(-90, -50),
figsize = size[0] / float(dpi), size[1] / float(dpi)
textcoords="offset points",
fig = plt.figure(figsize=figsize, dpi=dpi)
fontsize=16,
fig.patch.set_alpha(0)
arrowprops={"arrowstyle": "->", "connectionstyle": "arc3,rad=.2"},
plt.axes((0, 0.1, 1, 0.8), frameon=False)
)
for i in range(1, 11):
for label in ax.get_xticklabels() + ax.get_yticklabels():
plt.plot([i, i], [0, 1], color="b", lw=i / 2.0)
label.set_fontsize(16)
label.set_bbox({"facecolor": "white", "edgecolor": "None", "alpha": 0.65})
plt.xlim(0, 11)
plt.ylim(0, 1)
plt.show()
plt.xticks([])
plt.yticks([])
Total running time of the script: (0 minutes 0.097 seconds)
plt.show()

Total running time of the script: (0 minutes 0.014 seconds)

Example demoing choices for an option

The colors matplotlib line plots Alpha: transparency
An example demoing the various colors taken by matplotlib’s plot. This example demonstrates using alpha for transparency.

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

size = 256, 16 size = 256, 16
dpi = 72.0 dpi = 72.0
figsize = size[0] / float(dpi), size[1] / float(dpi) figsize = size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi) fig = plt.figure(figsize=figsize, dpi=dpi)
fig.patch.set_alpha(0) fig.patch.set_alpha(0)
plt.axes((0, 0.1, 1, 0.8), frameon=False) plt.axes((0, 0.1, 1, 0.8), frameon=False)
for i in range(1, 11): for i in range(1, 11):
plt.plot([i, i], [0, 1], lw=1.5) plt.axvline(i, linewidth=1, color="blue", alpha=0.25 + 0.75 * i / 10.0)
plt.xlim(0, 11) plt.xlim(0, 11)
plt.xticks([]) plt.xticks([])
plt.yticks([]) plt.yticks([])
plt.show() plt.show()

Total running time of the script: (0 minutes 0.013 seconds) Total running time of the script: (0 minutes 0.016 seconds)

4.7. Full code examples 170 4.7. Full code examples 171
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Aliased versus anti-aliased Marker size

This example demonstrates aliased versus anti-aliased text. Demo the marker size control in matplotlib.

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

size = 128, 16 size = 256, 16

dpi = 72.0 dpi = 72.0
figsize = size[0] / float(dpi), size[1] / float(dpi) figsize = size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi) fig = plt.figure(figsize=figsize, dpi=dpi)
fig.patch.set_alpha(0) fig.patch.set_alpha(0)
plt.axes((0, 0, 1, 1), frameon=False)
plt.axes((0, 0, 1, 1), frameon=False)
for i in range(1, 11):
plt.rcParams["text.antialiased"] = False plt.plot(
plt.text(0.5, 0.5, "Aliased", ha="center", va="center") [
i,
plt.xlim(0, 1) ],
plt.ylim(0, 1) [
plt.xticks([]) 1,
plt.yticks([]) ],
"s",
plt.show() markersize=i,
markerfacecolor="w",
Total running time of the script: (0 minutes 0.010 seconds) markeredgewidth=0.5,
markeredgecolor="k",
)

Aliased versus anti-aliased plt.xlim(0, 11)

plt.xticks([])
The example shows aliased versus anti-aliased text.
plt.yticks([])

plt.show()
import matplotlib.pyplot as plt
Total running time of the script: (0 minutes 0.014 seconds)
size = 128, 16
dpi = 72.0
figsize = size[0] / float(dpi), size[1] / float(dpi) Marker edge width
fig = plt.figure(figsize=figsize, dpi=dpi)
fig.patch.set_alpha(0) Demo the marker edge widths of matplotlib’s markers.
plt.axes((0, 0, 1, 1), frameon=False)

plt.rcParams["text.antialiased"] = True
plt.text(0.5, 0.5, "Anti-aliased", ha="center", va="center") import matplotlib.pyplot as plt

plt.xlim(0, 1) size = 256, 16

plt.ylim(0, 1) dpi = 72.0
plt.xticks([]) figsize = size[0] / float(dpi), size[1] / float(dpi)
plt.yticks([]) fig = plt.figure(figsize=figsize, dpi=dpi)
fig.patch.set_alpha(0)
plt.show() plt.axes((0, 0, 1, 1), frameon=False)

Total running time of the script: (0 minutes 0.010 seconds) for i in range(1, 11):
plt.plot(
[
(continues on next page)

4.7. Full code examples 172 4.7. Full code examples 173
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

i, maps.sort()
], l = len(maps) + 1
[
1, for i, m in enumerate(maps):
], plt.subplot(1, l, i + 1)
"s", plt.axis("off")
markersize=5, plt.imshow(a, aspect="auto", cmap=plt.get_cmap(m), origin="lower")
markeredgewidth=1 + i / 10.0, plt.title(m, rotation=90, fontsize=10, va="bottom")
markeredgecolor="k",
markerfacecolor="w", plt.show()
)
plt.xlim(0, 11) Total running time of the script: (0 minutes 1.792 seconds)
plt.xticks([])
plt.yticks([])

plt.show() Solid joint style

An example showing the different solid joint styles in matplotlib.

Total running time of the script: (0 minutes 0.014 seconds)

Colormaps import numpy as np

import matplotlib.pyplot as plt
An example plotting the matplotlib colormaps.
size = 256, 16
dpi = 72.0
figsize = size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi)
fig.patch.set_alpha(0)
plt.axes((0, 0, 1, 1), frameon=False)

plt.plot(np.arange(3), [0, 1, 0], color="blue", linewidth=8, solid_joinstyle="miter")

plt.plot(
4 + np.arange(3), [0, 1, 0], color="blue", linewidth=8, solid_joinstyle="bevel"
)
plt.plot(
8 + np.arange(3), [0, 1, 0], color="blue", linewidth=8, solid_joinstyle="round"
)

plt.xlim(0, 12)
plt.ylim(-1, 2)
plt.xticks([])
plt.yticks([])

plt.show()
import numpy as np
Total running time of the script: (0 minutes 0.011 seconds)
import matplotlib
import matplotlib.pyplot as plt

Solid cap style

plt.rc("text", usetex=False)
a = np.outer(np.arange(0, 1, 0.01), np.ones(10)) An example demoing the solide cap style in matplotlib.

plt.figure(figsize=(10, 5))
plt.subplots_adjust(top=0.8, bottom=0.05, left=0.01, right=0.99)
maps = [m for m in matplotlib.colormaps if not m.endswith("_r")]
(continues on next page)

4.7. Full code examples 174 4.7. Full code examples 175
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

import numpy as np
import matplotlib.pyplot as plt ],
[
size = 256, 16 1,
dpi = 72.0 ],
figsize = size[0] / float(dpi), size[1] / float(dpi) "s",
fig = plt.figure(figsize=figsize, dpi=dpi) markersize=5,
fig.patch.set_alpha(0) markerfacecolor="w",
plt.axes((0, 0, 1, 1), frameon=False) markeredgewidth=1.5,
markeredgecolor=(r, g, b, 1),
plt.plot(np.arange(4), np.ones(4), color="blue", linewidth=8, solid_capstyle="butt") )

plt.plot( plt.xlim(0, 11)

5 + np.arange(4), np.ones(4), color="blue", linewidth=8, solid_capstyle="round" plt.xticks([])
) plt.yticks([])

plt.plot( plt.show()
10 + np.arange(4),
np.ones(4), Total running time of the script: (0 minutes 0.015 seconds)
color="blue",
linewidth=8,
solid_capstyle="projecting",
Marker face color
)
Demo the marker face color of matplotlib’s markers.
plt.xlim(0, 14)
plt.xticks([])
plt.yticks([])
import numpy as np
plt.show() import matplotlib.pyplot as plt

Total running time of the script: (0 minutes 0.011 seconds) size = 256, 16
dpi = 72.0
figsize = size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi)
Marker edge color fig.patch.set_alpha(0)
plt.axes((0, 0, 1, 1), frameon=False)
Demo the marker edge color of matplotlib’s markers.
rng = np.random.default_rng()

import numpy as np for i in range(1, 11):

import matplotlib.pyplot as plt r, g, b = np.random.uniform(0, 1, 3)
plt.plot(
size = 256, 16 [
dpi = 72.0 i,
figsize = size[0] / float(dpi), size[1] / float(dpi) ],
fig = plt.figure(figsize=figsize, dpi=dpi) [
fig.patch.set_alpha(0) 1,
plt.axes((0, 0, 1, 1), frameon=False) ],
"s",
rng = np.random.default_rng() markersize=8,
markerfacecolor=(r, g, b, 1),
for i in range(1, 11): markeredgewidth=0.1,
r, g, b = np.random.uniform(0, 1, 3) markeredgecolor=(0, 0, 0, 0.5),
plt.plot( )
[ plt.xlim(0, 11)
i, plt.xticks([])
(continues on next page) (continues on next page)

4.7. Full code examples 176 4.7. Full code examples 177
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

plt.yticks([])
plt.show() Dash join style

Total running time of the script: (0 minutes 0.014 seconds) Example demoing the dash join style.

Dash capstyle import numpy as np

import matplotlib.pyplot as plt
An example demoing the dash capstyle.
size = 256, 16
dpi = 72.0
figsize = size[0] / float(dpi), size[1] / float(dpi)
import numpy as np fig = plt.figure(figsize=figsize, dpi=dpi)
import matplotlib.pyplot as plt fig.patch.set_alpha(0)
plt.axes((0, 0, 1, 1), frameon=False)
size = 256, 16
dpi = 72.0 plt.plot(
figsize = size[0] / float(dpi), size[1] / float(dpi) np.arange(3),
fig = plt.figure(figsize=figsize, dpi=dpi) [0, 1, 0],
fig.patch.set_alpha(0) color="blue",
plt.axes((0, 0, 1, 1), frameon=False) dashes=[12, 5],
linewidth=8,
plt.plot( dash_joinstyle="miter",
np.arange(4), )
np.ones(4), plt.plot(
color="blue", 4 + np.arange(3),
dashes=[15, 15], [0, 1, 0],
linewidth=8, color="blue",
dash_capstyle="butt", dashes=[12, 5],
) linewidth=8,
dash_joinstyle="bevel",
plt.plot( )
5 + np.arange(4), plt.plot(
np.ones(4), 8 + np.arange(3),
color="blue", [0, 1, 0],
dashes=[15, 15], color="blue",
linewidth=8, dashes=[12, 5],
dash_capstyle="round", linewidth=8,
) dash_joinstyle="round",
)
plt.plot(
10 + np.arange(4), plt.xlim(0, 12)
np.ones(4), plt.ylim(-1, 2)
color="blue", plt.xticks([])
dashes=[15, 15], plt.yticks([])
linewidth=8,
dash_capstyle="projecting", plt.show()
)
Total running time of the script: (0 minutes 0.011 seconds)
plt.xlim(0, 14)
plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.011 seconds)

4.7. Full code examples 178 4.7. Full code examples 179
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Markers (continued from previous page)

"d",
Show the different markers of matplotlib. ",",
"+",
"s",
"*",
"|",
"x",
"D",
"H",
".",
]

n_markers = len(markers)

size = 20 * n_markers, 300

dpi = 72.0
figsize = size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi)
plt.axes((0, 0.01, 1, 0.9), frameon=False)

for i, m in enumerate(markers):
import numpy as np marker(m, i)
import matplotlib.pyplot as plt
plt.xlim(-0.2, 0.2 + 0.5 * n_markers)
plt.xticks([])
def marker(m, i): plt.yticks([])
X = i * 0.5 * np.ones(11)
Y = np.arange(11) plt.show()

plt.plot(X, Y, lw=1, marker=m, ms=10, mfc=(0.75, 0.75, 1, 1), mec=(0, 0, 1, 1)) Total running time of the script: (0 minutes 0.060 seconds)
plt.text(0.5 * i, 10.25, repr(m), rotation=90, fontsize=15, va="bottom")

markers = [ Linestyles
0,
1, Plot the different line styles.
2,
3,
4,
5,
6,
7,
"o",
"h",
"_",
"1",
"2",
"3",
"4",
"8",
"p",
"^",
"v",
"<",
">",
"|",
(continues on next page)

4.7. Full code examples 180 4.7. Full code examples 181
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

"x",
"d",
"1",
"2",
"3",
"4",
"h",
"p",
"|",
"_",
"D",
"H",
]
n_lines = len(linestyles)

size = 20 * n_lines, 300

dpi = 72.0
figsize = size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi)
plt.axes((0, 0.01, 1, 0.9), frameon=False)

for i, ls in enumerate(linestyles):
import numpy as np linestyle(ls, i)
import matplotlib.pyplot as plt
plt.xlim(-0.2, 0.2 + 0.5 * n_lines)
plt.xticks([])
def linestyle(ls, i): plt.yticks([])
X = i * 0.5 * np.ones(11)
Y = np.arange(11) plt.show()
plt.plot(
X, Total running time of the script: (0 minutes 0.040 seconds)
Y,
ls,
color=(0.0, 0.0, 1, 1),
Locators for tick on axis
lw=3,
ms=8, An example demoing different locators to position ticks on axis for matplotlib.
mfc=(0.75, 0.75, 1, 1),
mec=(0, 0, 1, 1),
)
plt.text(0.5 * i, 10.25, ls, rotation=90, fontsize=15, va="bottom")

linestyles = [
"-",
"--",
":",
"-.",
".",
",",
"o",
"^",
"v",
"<",
">",
"s",
"+",
(continues on next page)

4.7. Full code examples 182 4.7. Full code examples 183
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

fig.patch.set_alpha(0)

for i, locator in enumerate(locators):

plt.subplot(n_locators, 1, i + 1)
ax = tickline()
ax.xaxis.set_major_locator(eval(locator))
plt.text(5, 0.3, locator[7:], ha="center")

plt.subplots_adjust(bottom=0.01, top=0.99, left=0.01, right=0.99)

plt.show()

Total running time of the script: (0 minutes 0.842 seconds)

Code generating the summary figures with a title

3D plotting vignette
import numpy as np Demo 3D plotting with matplotlib and decorate the figure.

from matplotlib import ticker

import matplotlib.pyplot as plt

def tickline():
plt.xlim(0, 10), plt.ylim(-1, 1), plt.yticks([])
ax = plt.gca()
ax.spines["right"].set_color("none")
ax.spines["left"].set_color("none")
ax.spines["top"].set_color("none")
ax.xaxis.set_ticks_position("bottom")
ax.spines["bottom"].set_position(("data", 0))
ax.yaxis.set_ticks_position("none")
ax.xaxis.set_minor_locator(ticker.MultipleLocator(0.1))
ax.plot(np.arange(11), np.zeros(11))
return ax

locators = [
"ticker.NullLocator()",
"ticker.MultipleLocator(1.0)",
"ticker.FixedLocator([0, 2, 8, 9, 10])",
"ticker.IndexLocator(3, 1)",
"ticker.LinearLocator(5)",
"ticker.LogLocator(2, [1.0])",
"ticker.AutoLocator()",
]

n_locators = len(locators)
import numpy as np
size = 512, 40 * n_locators import matplotlib.pyplot as plt
dpi = 72.0 from mpl_toolkits.mplot3d import Axes3D
figsize = size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi) x = np.arange(-4, 4, 0.25)
(continues on next page) (continues on next page)

4.7. Full code examples 184 4.7. Full code examples 185
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

y = np.arange(-4, 4, 0.25)
X, Y = np.meshgrid(x, y)
R = np.sqrt(X**2 + Y**2)
Z = np.sin(R)

fig = plt.figure()
ax: Axes3D = fig.add_subplot(111, projection="3d")

ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap="hot")

ax.contourf(X, Y, Z, zdir="z", offset=-2, cmap="hot")

ax.set_zlim(-2, 2)
plt.xticks([])
plt.yticks([])
ax.set_zticks([])

ax.text2D(
0.05,
0.93,
" 3D plots \n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
bbox={"facecolor": "white", "alpha": 1.0},
transform=plt.gca().transAxes,
)

ax.text2D(
0.05, import numpy as np
0.87,
" Plot 2D or 3D data", import matplotlib
horizontalalignment="left", import matplotlib.pyplot as plt
verticalalignment="top",
size="large",
transform=plt.gca().transAxes, plt.subplot(1, 1, 1, polar=True)
)
N = 20
plt.show() theta = np.arange(0.0, 2 * np.pi, 2 * np.pi / N)
rng = np.random.default_rng()
Total running time of the script: (0 minutes 0.110 seconds) radii = 10 * rng.random(N)
width = np.pi / 4 * rng.random(N)
bars = plt.bar(theta, radii, width=width, bottom=0.0)
jet = matplotlib.colormaps["jet"]
Plotting in polar, decorated
for r, bar in zip(radii, bars, strict=True):
An example showing how to plot in polar coordinate, and some decorations.
bar.set_facecolor(jet(r / 10.0))
bar.set_alpha(0.5)
plt.gca().set_xticklabels([])
plt.gca().set_yticklabels([])

plt.text(
-0.2,
1.02,
" Polar Axis \n",
horizontalalignment="left",
(continues on next page)

4.7. Full code examples 186 4.7. Full code examples 187
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

verticalalignment="top",
size="xx-large", n = 256
bbox={"facecolor": "white", "alpha": 1.0}, X = np.linspace(0, 2, n)
transform=plt.gca().transAxes, Y = np.sin(2 * np.pi * X)
)
plt.plot(X, Y, lw=2, color="violet")
plt.text( plt.xlim(-0.2, 2.2)
-0.2, plt.xticks([])
1.01, plt.ylim(-1.2, 1.2)
"\n\n Plot anything using polar axis ", plt.yticks([])
horizontalalignment="left",
verticalalignment="top",
size="large", # Add a title and a box around it
transform=plt.gca().transAxes, from matplotlib.patches import FancyBboxPatch
)
ax = plt.gca()
plt.show() ax.add_patch(
FancyBboxPatch(
Total running time of the script: (0 minutes 0.079 seconds) (-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
Plot example vignette boxstyle="square,pad=0",
zorder=3,
An example of plots with matplotlib, and added annotations.
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Regular Plot: plt.plot(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Plot lines and/or markers ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
)

plt.show()

Total running time of the script: (0 minutes 0.028 seconds)

import numpy as np
import matplotlib.pyplot as plt
(continues on next page)

4.7. Full code examples 188 4.7. Full code examples 189
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Multiple plots vignette (continued from previous page)

plt.text(
Demo multiple plots and style the figure. -0.05,
1.02,
" Multiplot: plt.subplot(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=ax.transAxes,
)
plt.text(
-0.05,
1.01,
"\n\n Plot several plots at once ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=ax.transAxes,
)

ax = plt.subplot(2, 2, 3)
ax.set_xticklabels([])
ax.set_yticklabels([])

ax = plt.subplot(2, 2, 4)
ax.set_xticklabels([])
ax.set_yticklabels([])

plt.show()

Total running time of the script: (0 minutes 0.077 seconds)

import matplotlib.pyplot as plt

Boxplot with matplotlib
ax = plt.subplot(2, 1, 1)
ax.set_xticklabels([]) An example of doing box plots with matplotlib
ax.set_yticklabels([])

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.72),
width=0.66,
height=0.34,
clip_on=False,
boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

(continues on next page)

4.7. Full code examples 190 4.7. Full code examples 191
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

)
)

plt.text(
-0.05,
1.02,
" Box Plot: plt.boxplot(...)\n ",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=axes.transAxes,
)

plt.text(
-0.04,
0.98,
"\n Make a box and whisker plot ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=axes.transAxes,
)

import numpy as np plt.show()

import matplotlib.pyplot as plt
Total running time of the script: (0 minutes 0.040 seconds)

fig = plt.figure(figsize=(8, 5))

axes = plt.subplot(111)
Plot scatter decorated
n = 5 An example showing the scatter function, with decorations.
Z = np.zeros((n, 4))
X = np.linspace(0, 2, n)
rng = np.random.default_rng()
Y = rng.random((n, 4))
plt.boxplot(Y)

plt.xticks([])
plt.yticks([])

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
(continues on next page)

4.7. Full code examples 192 4.7. Full code examples 193
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Scatter Plot: plt.scatter(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Make a scatter plot of x versus y ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
)

plt.show()
import numpy as np
import matplotlib.pyplot as plt Total running time of the script: (0 minutes 0.060 seconds)

n = 1024
rng = np.random.default_rng()
X = rng.normal(0, 1, n) Pie chart vignette
Y = rng.normal(0, 1, n)
Demo pie chart with matplotlib and style the figure.
T = np.arctan2(Y, X)

plt.scatter(X, Y, s=75, c=T, alpha=0.5)

plt.xlim(-1.5, 1.5)
plt.xticks([])
plt.ylim(-1.5, 1.5)
plt.yticks([])

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
boxstyle="square,pad=0",
(continues on next page)

4.7. Full code examples 194 4.7. Full code examples 195
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Pie Chart: plt.pie(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Make a pie chart of an array ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
)

import numpy as np plt.show()

import matplotlib.pyplot as plt
Total running time of the script: (0 minutes 0.052 seconds)
n = 20
X = np.ones(n)
X[-1] *= 2
plt.pie(X, explode=X * 0.05, colors=[f"{ i / float(n): f} " for i in range(n)]) Imshow demo

Demoing imshow
fig = plt.gcf()
w, h = fig.get_figwidth(), fig.get_figheight()
r = h / float(w)

plt.xlim(-1.5, 1.5)
plt.ylim(-1.5 * r, 1.5 * r)
plt.xticks([])
plt.yticks([])

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
(continues on next page)

4.7. Full code examples 196 4.7. Full code examples 197
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Imshow: plt.imshow(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Display an image to current axes ",
horizontalalignment="left",
verticalalignment="top",
family="DejaVu Sans",
size="large",
transform=plt.gca().transAxes,
)
import numpy as np
import matplotlib.pyplot as plt plt.show()

Total running time of the script: (0 minutes 0.035 seconds)

def f(x, y):
return (1 - x / 2 + x**5 + y**3) * np.exp(-(x**2) - y**2)

Bar plot advanced

n = 10
An more elaborate bar plot example
x = np.linspace(-3, 3, 8 * n)
y = np.linspace(-3, 3, 6 * n)
X, Y = np.meshgrid(x, y)
Z = f(X, Y)
plt.imshow(Z, interpolation="nearest", cmap="bone", origin="lower")
plt.xticks([])
plt.yticks([])

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
(continues on next page)

4.7. Full code examples 198 4.7. Full code examples 199
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Bar Plot: plt.bar(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Make a bar plot with rectangles ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
)

plt.show()

Total running time of the script: (0 minutes 0.045 seconds)

import numpy as np
import matplotlib.pyplot as plt

n = 16 Plotting quiver decorated

X = np.arange(n)
An example showing quiver with decorations.
Y1 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n)
Y2 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n)
plt.bar(X, Y1, facecolor="#9999ff", edgecolor="white")
plt.bar(X, -Y2, facecolor="#ff9999", edgecolor="white")
plt.xlim(-0.5, n)
plt.xticks([])
plt.ylim(-1, 1)
plt.yticks([])

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
boxstyle="square,pad=0",
zorder=3,
facecolor="white",
(continues on next page)

4.7. Full code examples 200 4.7. Full code examples 201
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Quiver Plot: plt.quiver(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Plot a 2-D field of arrows ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
)

import numpy as np
import matplotlib.pyplot as plt plt.show()

n = 8 Total running time of the script: (0 minutes 0.034 seconds)

X, Y = np.mgrid[0:n, 0:n]
T = np.arctan2(Y - n / 2.0, X - n / 2.0)
R = 10 + np.sqrt((Y - n / 2.0) ** 2 + (X - n / 2.0) ** 2)
U, V = R * np.cos(T), R * np.sin(T) Display the contours of a function

An example demoing how to plot the contours of a function, with additional layout tweaks.
plt.quiver(X, Y, U, V, R, alpha=0.5)
plt.quiver(X, Y, U, V, edgecolor="k", facecolor="None", linewidth=0.5)

plt.xlim(-1, n)
plt.xticks([])
plt.ylim(-1, n)
plt.yticks([])

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
(continues on next page)

4.7. Full code examples 202 4.7. Full code examples 203
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Contour Plot: plt.contour(..)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
/home/runner/work/scientific-python-lectures/scientific-python-lectures/intro/ )
˓→matplotlib/examples/pretty_plots/plot_contour_ext.py:24: UserWarning: The following␣

˓→kwargs were not used by contour: 'linewidth' plt.text(

C = plt.contour(X, Y, f(X, Y), 8, colors="black", linewidth=0.5) -0.05,
1.01,
"\n\n Draw contour lines and filled contours ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
import numpy as np )
import matplotlib.pyplot as plt
plt.show()

def f(x, y): Total running time of the script: (0 minutes 0.077 seconds)
return (1 - x / 2 + x**5 + y**3) * np.exp(-(x**2) - y**2)

n = 256 Grid elaborate

x = np.linspace(-3, 3, n)
y = np.linspace(-3, 3, n) An example displaying a grid on the axes and tweaking the layout.
X, Y = np.meshgrid(x, y)

plt.contourf(X, Y, f(X, Y), 8, alpha=0.75, cmap="hot")

C = plt.contour(X, Y, f(X, Y), 8, colors="black", linewidth=0.5)
plt.clabel(C, inline=1, fontsize=10)
plt.xticks([])
plt.yticks([])
(continues on next page)

4.7. Full code examples 204 4.7. Full code examples 205
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Grid: plt.grid(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=axes.transAxes,
)

plt.text(
Text(-0.05, 1.01, '\n\n Draw ticks and grid ') -0.05,
1.01,
"\n\n Draw ticks and grid ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=axes.transAxes,
import matplotlib.pyplot as plt )
from matplotlib.ticker import MultipleLocator
Total running time of the script: (0 minutes 0.095 seconds)
fig = plt.figure(figsize=(8, 6), dpi=72, facecolor="white")
axes = plt.subplot(111)
axes.set_xlim(0, 4)
axes.set_ylim(0, 3) Text printing decorated

axes.xaxis.set_major_locator(MultipleLocator(1.0)) An example showing text printing and decorating the resulting figure.
axes.xaxis.set_minor_locator(MultipleLocator(0.1))
axes.yaxis.set_major_locator(MultipleLocator(1.0))
axes.yaxis.set_minor_locator(MultipleLocator(0.1))
axes.grid(which="major", axis="x", linewidth=0.75, linestyle="-", color="0.75")
axes.grid(which="minor", axis="x", linewidth=0.25, linestyle="-", color="0.75")
axes.grid(which="major", axis="y", linewidth=0.75, linestyle="-", color="0.75")
axes.grid(which="minor", axis="y", linewidth=0.25, linestyle="-", color="0.75")
axes.set_xticklabels([])
axes.set_yticklabels([])

# Add a title and a box around it

(continues on next page)

4.7. Full code examples 206 4.7. Full code examples 207
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

x, y = rng.uniform(0, 1, 2)
alpha = rng.uniform(0.25, 0.75)
plt.text(
x,
y,
eq,
ha="center",
va="center",
color="#11557c",
alpha=alpha,
transform=plt.gca().transAxes,
fontsize=size,
clip_on=True,
)

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
boxstyle="square,pad=0",
zorder=3,
import numpy as np facecolor="white",
import matplotlib.pyplot as plt alpha=1.0,
transform=plt.gca().transAxes,
fig = plt.figure() )
plt.xticks([]) )
plt.yticks([])
plt.text(
eqs = [] -0.05,
eqs.append( 1.02,
r"$W^{3\beta}_{\delta_1 \rho_1 \sigma_2} = U^{3\beta}_{\delta_1 \rho_1} + \frac{1} " Text: plt.text(...)\n",
˓→ {8 \pi 2} \int^{\alpha_2}_{\alpha_2} d \alpha^\prime_2 \left[\frac{ U^{2\beta}_{\ horizontalalignment="left",
˓→delta_1 \rho_1} - \alpha^\prime_2U^{1\beta}_{\rho_1 \sigma_2} }{U^{0\beta}_{\rho_1 \ verticalalignment="top",
˓→sigma_2}}\right]$" size="xx-large",
) transform=plt.gca().transAxes,
eqs.append( )
r"$\frac{d\rho}{d t} + \rho \vec{v} \cdot\nabla\vec{v} = -\nabla p + \mu\nabla^2 \
˓→vec{v} + \rho \vec{g} $" plt.text(
) -0.05,
eqs.append(r"$\int_{-\infty}^\infty e^{-x^2}dx=\sqrt{\pi}$") 1.01,
eqs.append(r"$E = mc^2 = \sqrt{{m_0}^2c^4 + p^2c^2}$") "\n\n Draw any kind of text ",
eqs.append(r"$F_G = G\frac{m_1m_2} {r^2}$") horizontalalignment="left",
verticalalignment="top",
rng = np.random.default_rng() size="large",
transform=plt.gca().transAxes,
for i in range(24): )
index = rng.integers(0, len(eqs))
eq = eqs[index] plt.show()
size = rng.uniform(12, 32)
(continues on next page) Total running time of the script: (0 minutes 0.629 seconds)

4.7. Full code examples 208 4.7. Full code examples 209
 Scientific Python Lectures, Edition 2025.1rc0.dev0

CHAPTER 5
SciPy : high-level scientific computing

Authors: Gaël Varoquaux, Adrien Chauve, Andre Espaze, Emmanuelle Gouillart, Ralf Gommers

Scipy

The scipy package contains various toolboxes dedicated to common issues in scientific computing.
Its different submodules correspond to different applications, such as interpolation, integration, opti-
mization, image processing, statistics, special functions, etc.

 Tip

scipy can be compared to other standard scientific-computing libraries, such as the GSL (GNU
Scientific Library for C and C++), or Matlab’s toolboxes. scipy is the core package for scientific
routines in Python; it is meant to operate efficiently on numpy arrays, so that NumPy and SciPy work
hand in hand.
Before implementing a routine, it is worth checking if the desired data processing is not already
implemented in SciPy. As non-professional programmers, scientists often tend to re-invent the
wheel, which leads to buggy, non-optimal, difficult-to-share and unmaintainable code. By contrast,
SciPy’s routines are optimized and tested, and should therefore be used when possible.

Chapters contents

• File input/output: scipy.io

• Special functions: scipy.special

4.7. Full code examples 210 211

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• Linear algebra operations: scipy.linalg 5.1 File input/output: scipy.io

• Interpolation: scipy.interpolate scipy.io contains functions for loading and saving data in several common formats including Matlab,
IDL, Matrix Market, and Harwell-Boeing.
• Optimization and fit: scipy.optimize
Matlab files: Loading and saving:
• Statistics and random numbers: scipy.stats
• Numerical integration: scipy.integrate >>> import scipy as sp
>>> a = np.ones((3, 3))
• Fast Fourier transforms: scipy.fft >>> sp.io.savemat('file.mat', {'a': a}) # savemat expects a dictionary
• Signal processing: scipy.signal >>> data = sp.io.loadmat('file.mat')
>>> data['a']
• Image manipulation: scipy.ndimage array([[1., 1., 1.],
• Summary exercises on scientific computing [1., 1., 1.],
[1., 1., 1.]])
• Full code examples for the SciPy chapter

. Warning
. Warning Python / Matlab mismatch: The Matlab file format does not support 1D arrays.
This tutorial is far from an introduction to numerical computing. As enumerating the different >>> a = np.ones(3)
submodules and functions in SciPy would be very boring, we concentrate instead on a few examples >>> a
to give a general idea of how to use scipy for scientific computing. array([1., 1., 1.])
>>> a.shape
(3,)
scipy is composed of task-specific sub-modules: >>> sp.io.savemat('file.mat', {'a': a})
>>> a2 = sp.io.loadmat('file.mat')['a']
scipy.cluster Vector quantization / Kmeans >>> a2
scipy.constants Physical and mathematical constants array([[1., 1., 1.]])
scipy.fft Fourier transform >>> a2.shape
scipy.integrate Integration routines (1, 3)
scipy.interpolate Interpolation
Notice that the original array was a one-dimensional array, whereas the saved and reloaded array is
scipy.io Data input and output
a two-dimensional array with a single row.
scipy.linalg Linear algebra routines
scipy.ndimage n-dimensional image package For other formats, see the scipy.io documentation.
scipy.odr Orthogonal distance regression
scipy.optimize Optimization
scipy.signal Signal processing
scipy.sparse Sparse matrices ã See also
scipy.spatial Spatial data structures and algorithms
scipy.special Any special mathematical functions • Load text files: numpy.loadtxt()/numpy.savetxt()
scipy.stats Statistics • Clever loading of text/csv files: numpy.genfromtxt()
• Fast and efficient, but NumPy-specific, binary format: numpy.save()/numpy.load()
• Basic input/output of images in Matplotlib: matplotlib.pyplot.imread()/matplotlib.
 Tip pyplot.imsave()
They all depend on numpy, but are mostly independent of each other. The standard way of importing • More advanced input/output of images: imageio
NumPy and these SciPy modules is:
>>> import numpy as np
>>> import scipy as sp 5.2 Special functions: scipy.special
“Special” functions are functions commonly used in science and mathematics that are not considered to
be “elementary” functions. Examples include
• the gamma function, scipy.special.gamma(),
• the error function, scipy.special.erf(),

212 5.1. File input/output: scipy.io 213

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• Bessel functions, such as scipy.special.jv() (Bessel function of the first kind), and Both the numerator and denominator overflow, so performing 𝑎/𝑏 will not return the result we seek.
However, the magnitude of the result should be moderate, so the use of logarithms comes to mind.
• elliptic functions, such as scipy.special.ellipj() (Jacobi elliptic functions).
Combining the identities log(𝑎/𝑏) = log(𝑎) − log(𝑏) and exp(log(𝑥)) = 𝑥, we get:
Other special functions are combinations of familiar elementary functions, but they offer better accuracy
or robustness than their naive implementations would. >>> log_a = sp.special.gammaln(500)
>>> log_b = sp.special.gammaln(499)
Most of these function are computed elementwise and follow standard NumPy broadcasting rules when >>> log_res = log_a - log_b
the input arrays have different shapes. For example, scipy.special.xlog1py() is mathematically >>> res = np.exp(log_res)
equivalent to 𝑥 log(1 + 𝑦). >>> res
np.float64(499.0000000...)
>>> import scipy as sp
>>> x = np.asarray([1, 2])
>>> y = np.asarray([[3], [4], [5]]) Similarly, suppose we wish to compute the difference log(Γ(500) − Γ(499)). For this, we use scipy.
>>> res = sp.special.xlog1py(x, y) special.logsumexp(), which computes log(exp(𝑥)+exp(𝑦)) using a numerical trick that avoids overflow.
>>> res.shape >>> res = sp.special.logsumexp([log_a, log_b],
(3, 2) ... b=[1, -1]) # weights the terms of the sum
>>> ref = x * np.log(1 + y) >>> res
>>> np.allclose(res, ref) np.float64(2605.113844343...)
True
For more information about these and many other special functions, see the documentation of scipy.
However, scipy.special.xlog1py() is numerically favorable for small 𝑦, when explicit addition of 1 special.
would lead to loss of precision due to floating point truncation error.

>>> x = 2.5 5.3 Linear algebra operations: scipy.linalg

>>> y = 1e-18
>>> x * np.log(1 + y) scipy.linalg provides a Python interface to efficient, compiled implementations of standard linear alge-
np.float64(0.0) bra operations: the BLAS (Basic Linear Algebra Subroutines) and LAPACK (Linear Algebra PACKage)
>>> sp.special.xlog1py(x, y) libraries.
np.float64(2.5e-18)
For example, the scipy.linalg.det() function computes the determinant of a square matrix:
Many special functions also have “logarithmized” variants. For instance, the gamma function Γ(·) is
>>> import scipy as sp
related to the factorial function by 𝑛! = Γ(𝑛 + 1), but it extends the domain from the positive integers
>>> arr = np.array([[1, 2],
to the complex plane.
... [3, 4]])
>>> x = np.arange(10) >>> sp.linalg.det(arr)
>>> np.allclose(sp.special.gamma(x + 1), sp.special.factorial(x)) np.float64(-2.0)
True
>>> sp.special.gamma(5) < sp.special.gamma(5.5) < sp.special.gamma(6) Mathematically, the solution of a linear system 𝐴𝑥 = 𝑏 is 𝑥 = 𝐴−1 𝑏, but explicit inversion of a matrix is
np.True_ numerically unstable and should be avoided. Instead, use scipy.linalg.solve():

>>> A = np.array([[1, 2],

The factorial function grows quickly, and so the gamma function overflows for moderate values of the
... [2, 3]])
argument. However, sometimes only the logarithm of the gamma function is needed. In such cases, we
>>> b = np.array([14, 23])
can compute the logarithm of the gamma function directly using scipy.special.gammaln().
>>> x = sp.linalg.solve(A, b)
>>> x = [5, 50, 500] >>> x
>>> np.log(sp.special.gamma(x)) array([4., 5.])
array([ 3.17805383, 144.56574395, inf]) >>> np.allclose(A @ x, b)
>>> sp.special.gammaln(x) True
array([ 3.17805383, 144.56574395, 2605.11585036])
Linear systems with special structure can often be solved more efficiently than more general systems. For
Such functions can often be used when the intermediate components of a calculation would overflow example, systems with triangular matrices can be solved using scipy.linalg.solve_triangular():
or underflow, but the final result would not. For example, suppose we wish to compute the ratio
>>> A_upper = np.triu(A)
Γ(500)/Γ(499).
>>> A_upper
>>> a = sp.special.gamma(500) array([[1, 2],
>>> b = sp.special.gamma(499) [0, 3]])
>>> a, b >>> np.allclose(sp.linalg.solve_triangular(A_upper, b, lower=False),
(np.float64(inf), np.float64(inf)) ... sp.linalg.solve(A_upper, b))
True

5.2. Special functions: scipy.special 214 5.3. Linear algebra operations: scipy.linalg 215
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

scipy.linalg also features matrix factorizations/decompositions such as the singular value decomposi-
tion.

>>> A = np.array([[1, 2],

... [2, 3]])
>>> U, s, Vh = sp.linalg.svd(A)
>>> s # singular values
array([4.23606798, 0.23606798])

The original matrix can be recovered by matrix multiplication of the factors:

>>> S = np.diag(s) # convert to diagonal matrix before matrix multiplication

>>> A2 = U @ S @ Vh
>>> np.allclose(A2, A)
True
>>> A3 = (U * s) @ Vh # more efficient: use array math broadcasting rules!
On the other hand, if the data are not noisy, it may be desirable to pass exactly through each point.
>>> np.allclose(A3, A)
True >>> interp_spline = sp.interpolate.make_interp_spline(measured_time, function)
>>> interp_results = interp_spline(interpolation_time)
Many other decompositions (e.g. LU, Cholesky, QR), solvers for structured linear systems (e.g. triangu-
lar, circulant), eigenvalue problem algorithms, matrix functions (e.g. matrix exponential), and routines
for special matrix creation (e.g. block diagonal, toeplitz) are available in scipy.linalg.

5.4 Interpolation: scipy.interpolate

scipy.interpolate is used for fitting a function – an “interpolant” – to experimental or computed data.
Once fit, the interpolant can be used to approximate the underlying function at intermediate points; it
can also be used to compute the integral, derivative, or inverse of the function.
Some kinds of interpolants, known as “smoothing splines”, are designed to generate smooth curves from
noisy data. For example, suppose we have the following data:

>>> rng = np.random.default_rng(27446968)

>>> measured_time = np.linspace(0, 2*np.pi, 20)
>>> function = np.sin(measured_time)
>>> noise = rng.normal(loc=0, scale=0.1, size=20)
The derivative and antiderivative methods of the result object can be used for differentiation and
>>> measurements = function + noise
integration. For the latter, the constant of integration is assumed to be zero, but we can “wrap” the
antiderivative to include a nonzero constant of integration.
scipy.interpolate.make_smoothing_spline() can be used to form a curve similar to the underlying
sine function. >>> d_interp_spline = interp_spline.derivative()
>>> d_interp_results = d_interp_spline(interpolation_time)
>>> smoothing_spline = sp.interpolate.make_smoothing_spline(measured_time,␣ >>> i_interp_spline = lambda t: interp_spline.antiderivative()(t) - 1
˓→measurements)
>>> i_interp_results = i_interp_spline(interpolation_time)
>>> interpolation_time = np.linspace(0, 2*np.pi, 200)
>>> smooth_results = smoothing_spline(interpolation_time)

5.4. Interpolation: scipy.interpolate 216 5.4. Interpolation: scipy.interpolate 217

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

For functions that are monotonic on an interval (e.g. sin from 𝜋/2 to 3𝜋/2), we can reverse the arguments
>>> def f(x):
of make_interp_spline to interpolate the inverse function. Because the first argument is expected to
... # intersection of unit circle and line from origin
be monotonically increasing, we also reverse the order of elements in the arrays with numpy.flip().
... return [x[0]**2 + x[1]**2 - 1,
>>> i = (measured_time > np.pi/2) & (measured_time < 3*np.pi/2) ... x[1] - x[0]]
>>> inverse_spline = sp.interpolate.make_interp_spline(np.flip(function[i]), >>> res = sp.optimize.root(f, x0=[0, 0])
... np.flip(measured_time[i])) >>> np.allclose(f(res.x), 0, atol=1e-10)
>>> inverse_spline(0) True
array(3.14159265) >>> np.allclose(res.x, np.sqrt(2)/2)
True
See the summary exercise on Maximum wind speed prediction at the Sprogø station for a more advanced
spline interpolation example, and read the SciPy interpolation tutorial and the scipy.interpolate Over-constrained problems can be solved in the least-squares sense using scipy.optimize.root() with
documentation for much more information. method='lm' (Levenberg-Marquardt).

>>> def f(x):

5.5 Optimization and fit: scipy.optimize ... # intersection of unit circle, line from origin, and parabola
... return [x[0]**2 + x[1]**2 - 1,
scipy.optimize provides algorithms for root finding, curve fitting, and more general optimization. ... x[1] - x[0],
... x[1] - x[0]**2]
>>> res = sp.optimize.root(f, x0=[1, 1], method='lm')
5.5.1 Root Finding >>> res.success
scipy.optimize.root_scalar() attempts to find a root of a specified scalar-valued function (i.e., an True
argument at which the function value is zero). Like many scipy.optimize functions, the function needs >>> res.x
an initial guess of the solution, which the algorithm will refine until it converges or recognizes failure. array([0.76096066, 0.66017736])
We also provide the derivative to improve the rate of convergence.
See the documentation of scipy.optimize.root_scalar() and scipy.optimize.root() for a variety
>>> def f(x): of other solution algorithms and options.
... return (x-1)*(x-2)
>>> def df(x):
... return 2*x - 3 5.5.2 Curve fitting
>>> x0 = 0 # guess
>>> res = sp.optimize.root_scalar(f, x0=x0, fprime=df)
>>> res
converged: True
flag: converged
function_calls: 12
iterations: 6
root: 1.0
method: newton

. Warning

None of the functions in scipy.optimize that accept a guess are guaranteed to converge for all Suppose we have data that is sinusoidal but noisy:
possible guesses! (For example, try x0=1.5 in the example above, where the derivative of the function
is exactly zero.) If this occurs, try a different guess, adjust the options (like providing a bracket as >>> x = np.linspace(-5, 5, num=50) # 50 values between -5 and 5
shown below), or consider whether SciPy offers a more appropriate method for the problem. >>> noise = 0.01 * np.cos(100 * x)
>>> a, b = 2.9, 1.5
>>> y = a * np.cos(b * x) + noise
Note that only one the root at 1.0 is found. By inspection, we can tell that there is a second root at
2.0. We can direct the function toward a particular root by changing the guess or by passing a bracket We can approximate the underlying amplitude, frequency, and phase from the data by least squares
that contains only the root we seek. curve fitting. To begin, we write a function that accepts the independent variable as the first argument
and all parameters to fit as separate arguments:
>>> res = sp.optimize.root_scalar(f, bracket=(1.5, 10))
>>> res.root >>> def f(x, a, b, c):
2.0 ... return a * np.sin(b * x + c)

For multivariate problems, use scipy.optimize.root().

5.5. Optimization and fit: scipy.optimize 218 5.5. Optimization and fit: scipy.optimize 219
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Suppose we wish to minimize the scalar-valued function of a single variable 𝑓 (𝑥) = 𝑥2 + 10 sin(𝑥):

>>> def f(x):

... return x**2 + 10*np.sin(x)
>>> x = np.arange(-5, 5, 0.1)
>>> plt.plot(x, f(x))
[<matplotlib.lines.Line2D object at ...>]
>>> plt.show()

We can see that the function has a local minimizer near 𝑥 = 3.8 and a global minimizer near 𝑥 = −1.3,
but the precise values cannot be determined from the plot.
The most appropriate function for this purpose is scipy.optimize.minimize_scalar(). Since we know
the approximate locations of the minima, we will provide bounds that restrict the search to the vicinity
We then use scipy.optimize.curve_fit() to find 𝑎 and 𝑏:
of the global minimum.
>>> params, _ = sp.optimize.curve_fit(f, x, y, p0=[2, 1, 3])
>>> res = sp.optimize.minimize_scalar(f, bounds=(-2, -1))
>>> params
>>> res
array([2.900026 , 1.50012043, 1.57079633])
message: Solution found.
>>> ref = [a, b, np.pi/2] # what we'd expect
success: True
>>> np.allclose(params, ref, rtol=1e-3)
status: 0
True
fun: -7.9458233756...
x: -1.306440997...
nit: 8
Exercise: Curve fitting of temperature data nfev: 8
The temperature extremes in Alaska for each month, starting in January, are given by >>> res.fun == f(res.x)
(in degrees Celsius): np.True_

max: 17, 19, 21, 28, 33, 38, 37, 37, 31, 23, 19, 18 If we did not already know the approximate location of the global minimum, we could use one of
min: -62, -59, -56, -46, -32, -18, -9, -13, -25, -46, -52, -58 SciPy’s global minimizers, such as scipy.optimize.differential_evolution(). We are required to
pass bounds, but they do not need to be tight.
1. Plot these temperature extremes.
2. Define a function that can describe min and max temperatures. Hint: this function >>> bounds=[(-5, 5)] # list of lower, upper bound for each variable
has to have a period of 1 year. Hint: include a time offset. >>> res = sp.optimize.differential_evolution(f, bounds=bounds)
>>> res
3. Fit this function to the data with scipy.optimize.curve_fit(). message: Optimization terminated successfully.
4. Plot the result. Is the fit reasonable? If not, why? success: True
fun: -7.9458233756...
5. Is the time offset for min and max temperatures the same within the fit accuracy? x: [-1.306e+00]
solution nit: 6
nfev: 111
jac: [ 9.948e-06]
5.5.3 Optimization
For multivariate optimization, a good choice for many problems is scipy.optimize.minimize(). Sup-
pose we wish to find the minimum of a quadratic function of two variables, 𝑓 (𝑥0 , 𝑥1 ) = (𝑥0 −1)2 +(𝑥1 −2)2 .

>>> def f(x):

... return (x[0] - 1)**2 + (x[1] - 2)**2

Like scipy.optimize.root(), scipy.optimize.minimize() requires a guess x0. (Note that this is the
initial value of both variables rather than the value of the variable we happened to label 𝑥0 .)

>>> res = sp.optimize.minimize(f, x0=[0, 0])

>>> res
message: Optimization terminated successfully.
success: True
status: 0
fun: 1.70578...e-16
x: [ 1.000e+00 2.000e+00]
(continues on next page)

5.5. Optimization and fit: scipy.optimize 220 5.5. Optimization and fit: scipy.optimize 221
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) 5.6 Statistics and random numbers: scipy.stats
nit: 2
jac: [ 3.219e-09 -8.462e-09] scipy.stats contains fundamental tools for statistics in Python.
hess_inv: [[ 9.000e-01 -2.000e-01]
[-2.000e-01 6.000e-01]] 5.6.1 Statistical Distributions
nfev: 9
njev: 3 Consider a random variable distributed according to the standard normal. We draw a sample consisting
of 100000 observations from the random variable. The normalized histogram of the sample is an estimator
of the random variable’s probability density function (PDF):
Maximization? >>> dist = sp.stats.norm(loc=0, scale=1) # standard normal distribution
>>> sample = dist.rvs(size=100000) # "random variate sample"
Is scipy.optimize.minimize() restricted to the solution of minimization problems? Nope! To solve
>>> plt.hist(sample, bins=50, density=True, label='normalized histogram')
a maximization problem, simply minimize the negative of the original objective function.
>>> x = np.linspace(-5, 5)
>>> plt.plot(x, dist.pdf(x), label='PDF')
This barely scratches the surface of SciPy’s optimization features, which include mixed integer linear [<matplotlib.lines.Line2D object at ...>]
programming, constrained nonlinear programming, and the solution of assignment problems. For much >>> plt.legend()
more information, see the documentation of scipy.optimize and the advanced chapter Mathematical <matplotlib.legend.Legend object at ...>
optimization: finding minima of functions.

Exercise: 2-D minimization

The six-hump camelback function

𝑥4 2
𝑓 (𝑥, 𝑦) = (4 − 2.1𝑥2 + )𝑥 + 𝑥𝑦 + (4𝑦 2 − 4)𝑦 2
3
Distribution objects and frozen distributions
has multiple local minima. Find a global minimum (there is more than one, each with
the same value of the objective function) and at least one other local minimum. Each of the 100+ scipy.stats distribution families is represented by an object with a __call__
method. Here, we call the scipy.stats.norm object to specify its location and scale, and it returns
Hints:
a frozen distribution: a particular element of a distribution family with all parameters fixed. The
• Variables can be restricted to −2 < 𝑥 < 2 and −1 < 𝑦 < 1. frozen distribution object has methods to compute essential functions of the particular distribution.
• numpy.meshgrid() and matplotlib.pyplot.imshow() can help with visualization.
Suppose we knew that the sample had been drawn from a distribution belonging to the family of normal
• Try minimizing with scipy.optimize.minimize() with an initial guess of (𝑥, 𝑦) = distributions, but we did not know the particular distribution’s location (mean) and scale (standard
(0, 0). Does it find the global minimum, or converge to a local minimum? What deviation). We perform maximum likelihood estimation of the unknown parameters using the distribution
about other initial guesses? family’s fit method:
• Try minimizing with scipy.optimize.differential_evolution().
>>> loc, scale = sp.stats.norm.fit(sample)
solution >>> loc
np.float64(0.0015767005...)
See the summary exercise on Non linear least squares curve fitting: application to point extraction in >>> scale
topographical lidar data for another, more advanced example. np.float64(0.9973396878...)

5.5. Optimization and fit: scipy.optimize 222 5.6. Statistics and random numbers: scipy.stats 223
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Since we know the true parameters of the distribution from which the sample was drawn, we are not (continued from previous page)
surprised that these estimates are similar. >>> abs(integral - 1) < error_estimate # actual error < estimated error
True
Exercise: Probability distributions
Other functions for numerical quadrature, including integration of multivariate functions and approxi-
Generate 1000 random variates from a gamma distribution with a shape parameter of 1. Hint: the mating integrals from samples, are available in scipy.integrate.
shape parameter is passed as the first argument when freezing the distribution. Plot the histogram
of the sample, and overlay the distribution’s PDF. Estimate the shape parameter from the sample 5.7.2 Initial Value Problems
using the fit method.
scipy.integrate also features routines for integrating Ordinary Differential Equations (ODE). For
Extra: the distributions have many useful methods. Explore them using tab completion. Plot the example, scipy.integrate.solve_ivp() integrates ODEs of the form:
cumulative density function of the distribution, and compute the variance.
𝑑𝑦
= 𝑓 (𝑡, 𝑦(𝑡))
𝑑𝑡
5.6.2 Sample Statistics and Hypothesis Tests
from an initial time 𝑡0 and initial state 𝑦(𝑡 = 𝑡0 ) = 𝑡0 to a final time 𝑡𝑓 or until an event occurs (e.g. a
The sample mean is an estimator of the mean of the distribution from which the sample was drawn: specified state is reached).
>>> np.mean(sample) As an introduction, consider the initial value problem given by 𝑑𝑦 𝑑𝑡 = −2𝑦 and the initial condition
np.float64(0.001576700508...) 𝑦(𝑡 = 0) = 1 on the interval 𝑡 = 0 . . . 4. We begin by defining a callable that computes 𝑓 (𝑡, 𝑦(𝑡)) given
the current time and state.
NumPy includes some of the most fundamental sample statistics (e.g. numpy.mean(), numpy.var(),
>>> def f(t, y):
numpy.percentile()); scipy.stats includes many more. For instance, the geometric mean is a common
... return -2 * y
measure of central tendency for data that tends to be distributed over many orders of magnitude.

>>> sp.stats.gmean(2**sample) Then, to compute y as a function of time:

np.float64(1.0010934829...)
>>> t_span = (0, 4) # time interval
>>> t_eval = np.linspace(*t_span) # times at which to evaluate `y`
SciPy also includes a variety of hypothesis tests that produce a sample statistic and a p-value. For
>>> y0 = [1,] # initial state
instance, suppose we wish to test the null hypothesis that sample was drawn from a normal distribution:
>>> res = sp.integrate.solve_ivp(f, t_span=t_span, y0=y0, t_eval=t_eval)
>>> res = sp.stats.normaltest(sample)
>>> res.statistic and plot the result:
np.float64(5.20841759...)
>>> plt.plot(res.t, res.y[0])
>>> res.pvalue
[<matplotlib.lines.Line2D object at ...>]
np.float64(0.07396163283...)
>>> plt.xlabel('t')
Text(0.5, ..., 't')
Here, statistic is a sample statistic that tends to be high for samples that are drawn from non-normal
>>> plt.ylabel('y')
distributions. pvalue is the probability of observing such a high value of the statistic for a sample
Text(..., 0.5, 'y')
that has been drawn from a normal distribution. If the p-value is unusually small, this may be taken
>>> plt.title('Solution of Initial Value Problem')
as evidence that sample was not drawn from the normal distribution. Our statistic and p-value are
Text(0.5, 1.0, 'Solution of Initial Value Problem')
moderate, so the test is inconclusive.
There are many other features of scipy.stats, including circular statistics, quasi-Monte Carlo methods,
and resampling methods. For much more information, see the documentation of scipy.stats and the
advanced chapter statistics.

5.7 Numerical integration: scipy.integrate

5.7.1 Quadrature
∫︀ 𝜋/2
Suppose we wish to compute the definite integral 0 sin(𝑡)𝑑𝑡 numerically. scipy.integrate.quad()
chooses one of several adaptive techniques depending on the parameters, and is therefore the recom-
mended first choice for integration of function of a single variable:

>>> integral, error_estimate = sp.integrate.quad(np.sin, 0, np.pi/2)

>>> np.allclose(integral, 1) # numerical result ~ analytical result Let us integrate a more complex ODE: a damped spring-mass oscillator. The position of√︀ a mass attached
True to a spring obeys the 2nd order ODE 𝑦¨ + 2𝜁𝜔0 𝑦˙ + 𝜔02 𝑦 = 0 with natural frequency 𝜔0 = 𝑘/𝑚, damping
(continues on next page) ratio 𝜁 = 𝑐/(2𝑚𝜔0 ), spring constant 𝑘, mass 𝑚, and damping coefficient 𝑐.

5.7. Numerical integration: scipy.integrate 224 5.7. Numerical integration: scipy.integrate 225
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Before using scipy.integrate.solve_ivp(), the 2nd order ODE needs to be transformed into a system 5.8 Fast Fourier transforms: scipy.fft
of first-order ODEs. Note that
𝑑𝑦 𝑑𝑦˙ The scipy.fft module computes fast Fourier transforms (FFTs) and offers utilities to handle them.
= 𝑦˙ = 𝑦¨ = −(2𝜁𝜔0 𝑦˙ + 𝜔02 𝑦)
𝑑𝑡 𝑑𝑡 Some important functions are:
If we define 𝑧 = [𝑧0 , 𝑧1 ] where 𝑧0 = 𝑦 and 𝑧1 = 𝑦,˙ then the first order equation: • scipy.fft.fft() to compute the FFT
[︂ 𝑑𝑧0 ]︂ [︂ ]︂
𝑑𝑧 𝑧1 • scipy.fft.fftfreq() to generate the sampling frequencies
= 𝑑𝑧 𝑑𝑡 = 2
𝑑𝑡 1
−(2𝜁𝜔 𝑧
0 1 + 𝜔 𝑧
0 0 )
𝑑𝑡 • scipy.fft.ifft() to compute the inverse FFT, from frequency space to signal space
is equivalent to the original second order equation.
We set:

>>> m = 0.5 # kg
>>> k = 4 # N/m
>>> c = 0.4 # N s/m As an illustration, a (noisy) input signal (sig), and its FFT:
>>> zeta = c / (2 * m * np.sqrt(k/m)) >>> sig_fft = sp.fft.fft(sig)
>>> omega = np.sqrt(k / m) >>> freqs = sp.fft.fftfreq(sig.size, d=time_step)

and define the function that computes 𝑧˙ = 𝑓 (𝑡, 𝑧(𝑡)):

>>> def f(t, z, zeta, omega):

... return (z[1], -2.0 * zeta * omega * z[1] - omega**2 * z[0])

Signal FFT
Integration of the system follows:

>>> t_span = (0, 10) As the signal comes from a real-valued function, the Fourier transform is symmetric.
>>> t_eval = np.linspace(*t_span, 100) The peak signal frequency can be found with freqs[power.argmax()]
>>> z0 = [1, 0]
>>> res = sp.integrate.solve_ivp(f, t_span, z0, t_eval=t_eval,
... args=(zeta, omega), method='LSODA')

 Tip

With the option method=’LSODA’, scipy.integrate.solve_ivp() uses the LSODA (Livermore

Solver for Ordinary Differential equations with Automatic method switching for stiff and non-stiff
problems). See the ODEPACK Fortran library for more details.

ã See also

Partial Differental Equations

There is no Partial Differential Equations (PDE) solver in SciPy. Some Python packages for solving
PDE’s are available, such as fipy or SfePy.

5.7. Numerical integration: scipy.integrate 226 5.8. Fast Fourier transforms: scipy.fft 227
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Exercise: Denoise moon landing image

Setting the Fourier component above this frequency to zero and inverting the FFT with scipy.fft.
ifft(), gives a filtered signal.

ò Note

The code of this example can be found here

numpy.fft

NumPy also has an implementation of FFT (numpy.fft). However, the SciPy one should be preferred,
as it uses more efficient underlying implementations.
1. Examine the provided image moonlanding.png, which is heavily contaminated with periodic
noise. In this exercise, we aim to clean up the noise using the Fast Fourier Transform.
2. Load the image using matplotlib.pyplot.imread().
3. Find and use the 2-D FFT function in scipy.fft, and plot the spectrum (Fourier transform
of) the image. Do you have any trouble visualising the spectrum? If so, why?
Fully worked examples:
4. The spectrum consists of high and low frequency components. The noise is contained in the
high-frequency part of the spectrum, so set some of those components to zero (use array slicing).
Crude periodicity finding (link) Gaussian image blur (link)
5. Apply the inverse Fourier transform to see the resulting image.
Solution

5.9 Signal processing: scipy.signal

 Tip

scipy.signal is for typical signal processing: 1D, regularly-sampled signals.

5.8. Fast Fourier transforms: scipy.fft 228 5.9. Signal processing: scipy.signal 229
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> x_detrended = sp.signal.detrend(x)

>>> plt.plot(t, x)
[<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(t, x_detrended)
[<matplotlib.lines.Line2D object at ...>]

Filtering: For non-linear filtering, scipy.signal has filtering (median filter scipy.signal.medfilt(),
Wiener scipy.signal.wiener()), but we will discuss this in the image section.

 Tip

scipy.signal also has a full-blown set of tools for the design of linear filter (finite and infinite
response filters), but this is out of the scope of this tutorial.
Resampling scipy.signal.resample(): resample a signal to n points using FFT.

>>> t = np.linspace(0, 5, 100) Spectral analysis: scipy.signal.spectrogram() compute a spectrogram –frequency spectrums over
>>> x = np.sin(t) consecutive time windows–, while scipy.signal.welch() comptes a power spectrum density (PSD).

>>> x_resampled = sp.signal.resample(x, 25)

>>> plt.plot(t, x)
[<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(t[::4], x_resampled, 'ko')
[<matplotlib.lines.Line2D object at ...>]

 Tip

Notice how on the side of the window the resampling is less accurate and has a rippling effect.
This resampling is different from the interpolation provided by scipy.interpolate as it only applies
to regularly sampled data.

5.10 Image manipulation: scipy.ndimage

scipy.ndimage provides manipulation of n-dimensional arrays as images.

5.10.1 Geometrical transformations on images

Detrending scipy.signal.detrend(): remove linear trend from signal:
Changing orientation, resolution, ..
>>> t = np.linspace(0, 5, 100)
>>> rng = np.random.default_rng() >>> import scipy as sp
>>> x = t + rng.normal(size=100)
(continues on next page) (continues on next page)

5.9. Signal processing: scipy.signal 230 5.10. Image manipulation: scipy.ndimage 231
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> # Load an image
>>> face = sp.datasets.face(gray=True)

>>> # Shift, rotate and zoom it

>>> shifted_face = sp.ndimage.shift(face, (50, 50))
>>> shifted_face2 = sp.ndimage.shift(face, (50, 50), mode='nearest')
>>> rotated_face = sp.ndimage.rotate(face, 30)
>>> cropped_face = face[50:-50, 50:-50]
>>> zoomed_face = sp.ndimage.zoom(face, 2)
>>> zoomed_face.shape
(1536, 2048)
Other filters in scipy.ndimage.filters and scipy.signal can be applied to images.

Exercise

Compare histograms for the different filtered images.

5.10.3 Mathematical morphology

>>> plt.subplot(151)  Tip

<Axes: >
Mathematical morphology stems from set theory. It characterizes and transforms geometrical struc-
>>> plt.imshow(shifted_face, cmap=plt.cm.gray) tures. Binary (black and white) images, in particular, can be transformed using this theory: the sets
<matplotlib.image.AxesImage object at 0x...> to be transformed are the sets of neighboring non-zero-valued pixels. The theory was also extended
to gray-valued images.
>>> plt.axis('off')
(np.float64(-0.5), np.float64(1023.5), np.float64(767.5), np.float64(-0.5))

>>> # etc.

5.10.2 Image filtering

Generate a noisy face:

>>> import scipy as sp

>>> face = sp.datasets.face(gray=True)
>>> face = face[:512, -512:] # crop out square on right
>>> import numpy as np
>>> noisy_face = np.copy(face).astype(float)
>>> rng = np.random.default_rng() Mathematical-morphology operations use a structuring element in order to modify geometrical structures.
>>> noisy_face += face.std() * 0.5 * rng.standard_normal(face.shape) Let us first generate a structuring element:

Apply a variety of filters on it: >>> el = sp.ndimage.generate_binary_structure(2, 1)

>>> el
>>> blurred_face = sp.ndimage.gaussian_filter(noisy_face, sigma=3) array([[False, True, False],
>>> median_face = sp.ndimage.median_filter(noisy_face, size=5) [...True, True, True],
>>> wiener_face = sp.signal.wiener(noisy_face, (5, 5)) [False, True, False]])
>>> el.astype(int)
array([[0, 1, 0],
[1, 1, 1],
[0, 1, 0]])

• Erosion scipy.ndimage.binary_erosion()

5.10. Image manipulation: scipy.ndimage 232 5.10. Image manipulation: scipy.ndimage 233
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> a = np.zeros((7, 7), dtype=int)
>>> a[1:6, 2:5] = 1 >>> sp.ndimage.binary_opening(a, structure=np.ones((3, 3))).astype(int)
>>> a array([[0, 0, 0, 0, 0],
array([[0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 1, 0],
[0, 0, 1, 1, 1, 0, 0], [0, 1, 1, 1, 0],
[0, 0, 1, 1, 1, 0, 0], [0, 1, 1, 1, 0],
[0, 0, 1, 1, 1, 0, 0], [0, 0, 0, 0, 0]])
[0, 0, 1, 1, 1, 0, 0], >>> # Opening can also smooth corners
[0, 0, 1, 1, 1, 0, 0], >>> sp.ndimage.binary_opening(a).astype(int)
[0, 0, 0, 0, 0, 0, 0]]) array([[0, 0, 0, 0, 0],
>>> sp.ndimage.binary_erosion(a).astype(a.dtype) [0, 0, 1, 0, 0],
array([[0, 0, 0, 0, 0, 0, 0], [0, 1, 1, 1, 0],
[0, 0, 0, 0, 0, 0, 0], [0, 0, 1, 0, 0],
[0, 0, 0, 1, 0, 0, 0], [0, 0, 0, 0, 0]])
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0], • Closing: scipy.ndimage.binary_closing()
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
Exercise
>>> # Erosion removes objects smaller than the structure
>>> sp.ndimage.binary_erosion(a, structure=np.ones((5,5))).astype(a.dtype) Check that opening amounts to eroding, then dilating.
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0], An opening operation removes small structures, while a closing operation fills small holes. Such opera-
[0, 0, 0, 0, 0, 0, 0], tions can therefore be used to “clean” an image.
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0], >>> a = np.zeros((50, 50))
[0, 0, 0, 0, 0, 0, 0], >>> a[10:-10, 10:-10] = 1
[0, 0, 0, 0, 0, 0, 0]]) >>> rng = np.random.default_rng()
>>> a += 0.25 * rng.standard_normal(a.shape)
• Dilation scipy.ndimage.binary_dilation() >>> mask = a>=0.5
>>> opened_mask = sp.ndimage.binary_opening(mask)
>>> a = np.zeros((5, 5)) >>> closed_mask = sp.ndimage.binary_closing(opened_mask)
>>> a[2, 2] = 1
>>> a
array([[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 1., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.]])
>>> sp.ndimage.binary_dilation(a).astype(a.dtype)
array([[0., 0., 0., 0., 0.],
[0., 0., 1., 0., 0.],
[0., 1., 1., 1., 0.],
[0., 0., 1., 0., 0.],
[0., 0., 0., 0., 0.]])

• Opening scipy.ndimage.binary_opening() Exercise

>>> a = np.zeros((5, 5), dtype=int) Check that the area of the reconstructed square is smaller than the area of the initial
>>> a[1:4, 1:4] = 1 square. (The opposite would occur if the closing step was performed before the opening).
>>> a[4, 4] = 1
>>> a For gray-valued images, eroding (resp. dilating) amounts to replacing a pixel by the minimal (resp.
array([[0, 0, 0, 0, 0], maximal) value among pixels covered by the structuring element centered on the pixel of interest.
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0], >>> a = np.zeros((7, 7), dtype=int)
[0, 1, 1, 1, 0], >>> a[1:6, 1:6] = 3
[0, 0, 0, 0, 1]]) >>> a[4, 4] = 2; a[2, 3] = 1
>>> # Opening removes small objects >>> a
(continues on next page) (continues on next page)

5.10. Image manipulation: scipy.ndimage 234 5.10. Image manipulation: scipy.ndimage 235
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> areas = sp.ndimage.sum(mask, labels, range(1, labels.max()+1))
array([[0, 0, 0, 0, 0, 0, 0], >>> areas # The number of pixels in each connected component
[0, 3, 3, 3, 3, 3, 0], array([190., 45., 424., 278., 459., 190., 549., 424.])
[0, 3, 3, 1, 3, 3, 0], >>> maxima = sp.ndimage.maximum(sig, labels, range(1, labels.max()+1))
[0, 3, 3, 3, 3, 3, 0], >>> maxima # The maximum signal in each connected component
[0, 3, 3, 3, 2, 3, 0], array([ 1.80238238, 1.13527605, 5.51954079, 2.49611818, 6.71673619,
[0, 3, 3, 3, 3, 3, 0], 1.80238238, 16.76547217, 5.51954079])
[0, 0, 0, 0, 0, 0, 0]])
>>> sp.ndimage.grey_erosion(a, size=(3, 3))
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 3, 2, 2, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])

5.10.4 Connected components and measurements on images

Let us first generate a nice synthetic binary image.
>>> x, y = np.indices((100, 100))
Extract the 4th connected component, and crop the array around it:
>>> sig = np.sin(2*np.pi*x/50.) * np.sin(2*np.pi*y/50.) * (1+x*y/50.**2)**2
>>> mask = sig > 1 >>> sp.ndimage.find_objects(labels)[3]
(slice(30, 48, None), slice(30, 48, None))
>>> sl = sp.ndimage.find_objects(labels)[3]
>>> import matplotlib.pyplot as plt
>>> plt.imshow(sig[sl])
<matplotlib.image.AxesImage object at ...>

See the summary exercise on Image processing application: counting bubbles and unmolten grains for a
more advanced example.

5.11 Summary exercises on scientific computing

The summary exercises use mainly NumPy, SciPy and Matplotlib. They provide some real-life examples
of scientific computing with Python. Now that the basics of working with NumPy and SciPy have been
introduced, the interested user is invited to try these exercises.

5.11.1 Maximum wind speed prediction at the Sprogø station

The exercise goal is to predict the maximum wind speed occurring every 50 years even if no measure
exists for such a period. The available data are only measured over 21 years at the Sprogø meteorological
station located in Denmark. First, the statistical steps will be given and then illustrated with functions
from the scipy.interpolate module. At the end the interested readers are invited to compute results from
raw data and in a slightly different approach.

Statistical approach
The annual maxima are supposed to fit a normal probability density function. However such function
is not going to be estimated because it gives a probability from a wind speed maxima. Finding the
scipy.ndimage.label() assigns a different label to each connected component: maximum wind speed occurring every 50 years requires the opposite approach, the result needs to be
found from a defined probability. That is the quantile function role and the exercise goal will be to find
>>> labels, nb = sp.ndimage.label(mask)
it. In the current model, it is supposed that the maximum wind speed occurring every 50 years is defined
>>> nb
as the upper 2% quantile.
8
By definition, the quantile function is the inverse of the cumulative distribution function. The latter
Now compute measurements on each connected component: describes the probability distribution of an annual maxima. In the exercise, the cumulative probability

5.10. Image manipulation: scipy.ndimage 236 5.11. Summary exercises on scientific computing 237
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

p_i for a given year i is defined as p_i = i/(N+1) with N = 21, the number of measured years. Thus
it will be possible to calculate the cumulative probability of every measured wind speed maxima. From
those experimental points, the scipy.interpolate module will be very useful for fitting the quantile function.
Finally the 50 years maxima is going to be evaluated from the cumulative probability of the 2% quantile.

Computing the cumulative probabilities

The annual wind speeds maxima have already been computed and saved in the NumPy format in the
file examples/max-speeds.npy, thus they will be loaded by using NumPy:

>>> import numpy as np

>>> max_speeds = np.load('intro/scipy/summary-exercises/examples/max-speeds.npy')
>>> years_nb = max_speeds.shape[0]

Following the cumulative probability definition p_i from the previous section, the corresponding values
will be:

>>> cprob = (np.arange(years_nb, dtype=np.float32) + 1)/(years_nb + 1)

and they are assumed to fit the given wind speeds:

>>> sorted_max_speeds = np.sort(max_speeds)

Prediction with UnivariateSpline

In this section the quantile function will be estimated by using the UnivariateSpline class which can
represent a spline from points. The default behavior is to build a spline of degree 3 and points can
have different weights according to their reliability. Variants are InterpolatedUnivariateSpline and
LSQUnivariateSpline on which errors checking is going to change. In case a 2D spline is wanted, the
BivariateSpline class family is provided. All those classes for 1D and 2D splines use the FITPACK
Fortran subroutines, that’s why a lower library access is available through the splrep and splev functions
for respectively representing and evaluating a spline. Moreover interpolation functions without the use
of FITPACK parameters are also provided for simpler use.
For the Sprogø maxima wind speeds, the UnivariateSpline will be used because a spline of degree 3
seems to correctly fit the data:

>>> import scipy as sp

>>> quantile_func = sp.interpolate.UnivariateSpline(cprob, sorted_max_speeds)

The quantile function is now going to be evaluated from the full range of probabilities:

>>> nprob = np.linspace(0, 1, 100)

>>> fitted_max_speeds = quantile_func(nprob)

In the current model, the maximum wind speed occurring every 50 years is defined as the upper 2% Fig. 1: Solution: Python source file
quantile. As a result, the cumulative probability value will be:

>>> fifty_prob = 1. - 0.02

So the storm wind speed occurring every 50 years can be guessed by:

>>> fifty_wind = quantile_func(fifty_prob)

>>> fifty_wind
array(32.97989825...)

The results are now gathered on a Matplotlib figure:

5.11. Summary exercises on scientific computing 238 5.11. Summary exercises on scientific computing 239
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Exercise with the Gumbell distribution

The interested readers are now invited to make an exercise by using the wind speeds measured over 21
years. The measurement period is around 90 minutes (the original period was around 10 minutes but the
file size has been reduced for making the exercise setup easier). The data are stored in NumPy format
inside the file examples/sprog-windspeeds.npy. Do not look at the source code for the plots until you
have completed the exercise.
• The first step will be to find the annual maxima by using NumPy and plot them as a matplotlib
bar figure.

Fig. 2: Solution: Python source file

• The second step will be to use the Gumbell distribution on cumulative probabilities p_i defined as
-log( -log(p_i) ) for fitting a linear quantile function (remember that you can define the degree
of the UnivariateSpline). Plotting the annual maxima versus the Gumbell distribution should Fig. 3: Solution: Python source file
give you the following figure.
• The last step will be to find 34.23 m/s for the maximum wind speed occurring every 50 years.

5.11.2 Non linear least squares curve fitting: application to point extraction in
topographical lidar data
The goal of this exercise is to fit a model to some data. The data used in this tutorial are lidar data
and are described in details in the following introductory paragraph. If you’re impatient and want to
practice now, please skip it and go directly to Loading and visualization.

5.11. Summary exercises on scientific computing 240 5.11. Summary exercises on scientific computing 241
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Introduction
Lidars systems are optical rangefinders that analyze property of scattered light to measure distances.
Most of them emit a short light impulsion towards a target and record the reflected signal. This signal
is then processed to extract the distance between the lidar system and the target.
Topographical lidar systems are such systems embedded in airborne platforms. They measure distances
between the platform and the Earth, so as to deliver information on the Earth’s topography (see1 for
more details).
In this tutorial, the goal is to analyze the waveform recorded by the lidar system2 . Such a signal
contains peaks whose center and amplitude permit to compute the position and some characteristics of
the hit target. When the footprint of the laser beam is around 1m on the Earth surface, the beam can
hit multiple targets during the two-way propagation (for example the ground and the top of a tree or
building). The sum of the contributions of each target hit by the laser beam then produces a complex
signal with multiple peaks, each one containing information about one target.
One state of the art method to extract information from these data is to decompose them in a sum of
Gaussian functions where each function represents the contribution of a target hit by the laser beam.
Therefore, we use the scipy.optimize module to fit a waveform to one or a sum of Gaussian functions.

Loading and visualization

Load the first waveform using:

>>> import numpy as np

>>> waveform_1 = np.load('intro/scipy/summary-exercises/examples/waveform_1.npy')

and visualize it:

>>> import matplotlib.pyplot as plt

>>> t = np.arange(len(waveform_1))
>>> plt.plot(t, waveform_1)
[<matplotlib.lines.Line2D object at ...>] >>> def model(t, coeffs):
>>> plt.show() ... return coeffs[0] + coeffs[1] * np.exp( - ((t-coeffs[2])/coeffs[3])**2 )

As shown below, this waveform is a 80-bin-length signal with a single peak with an amplitude of ap- where
proximately 30 in the 15 nanosecond bin. Additionally, the base level of noise is approximately 3. These • coeffs[0] is 𝐵 (noise)
values can be used in the initial solution.
• coeffs[1] is 𝐴 (amplitude)
Fitting a waveform with a simple Gaussian model • coeffs[2] is 𝜇 (center)
The signal is very simple and can be modeled as a single Gaussian function and an offset corresponding • coeffs[3] is 𝜎 (width)
to the background noise. To fit the signal with the function, we must:
• define the model Initial solution

• propose an initial solution One possible initial solution that we determine by inspection is:
• call scipy.optimize.leastsq >>> x0 = np.array([3, 30, 15, 1], dtype=float)

Model
Fit
A Gaussian function defined by
{︃ (︂ scipy.optimize.leastsq minimizes the sum of squares of the function given as an argument. Basically,
)︂2 }︃
𝑡−𝜇 the function to minimize is the residuals (the difference between the data and the model):
𝐵 + 𝐴 exp −
𝜎
>>> def residuals(coeffs, y, t):
can be defined in python by: ... return y - model(t, coeffs)
1 Mallet, C. and Bretar, F. Full-Waveform Topographic Lidar: State-of-the-Art. ISPRS Journal of Photogrammetry

and Remote Sensing 64(1), pp.1-16, January 2009 http://dx.doi.org/10.1016/j.isprsjprs.2008.09.007 So let’s get our solution by calling scipy.optimize.leastsq() with the following arguments:
2 The data used for this tutorial are part of the demonstration data available for the FullAnalyze software and were

kindly provided by the GIS DRAIX.

• the function to minimize

5.11. Summary exercises on scientific computing 242 5.11. Summary exercises on scientific computing 243
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• an initial solution
• the additional arguments to pass to the function

>>> import scipy as sp

>>> t = np.arange(len(waveform_1))
>>> x, flag = sp.optimize.leastsq(residuals, x0, args=(waveform_1, t))
>>> x
array([ 2.70363, 27.82020, 15.47924, 3.05636])

And visualize the solution:

fig, ax = plt.subplots(figsize=(8, 6))

plt.plot(t, waveform_1, t, model(t, x))
plt.xlabel("Time [ns]")
plt.ylabel("Amplitude [bins]")
plt.legend(["Waveform", "Model"])
plt.show()

an input for leastsq.

• When we want to detect very small peaks in the signal, or when the initial guess is too far from a
good solution, the result given by the algorithm is often not satisfying. Adding constraints to the
parameters of the model enables to overcome such limitations. An example of a priori knowledge
we can add is the sign of our variables (which are all positive).
• See the solution.
• Further exercise: compare the result of scipy.optimize.leastsq() and what you can get with
scipy.optimize.fmin_slsqp() when adding boundary constraints.

Remark: from scipy v0.8 and above, you should rather use scipy.optimize.curve_fit() which takes
the model and the data as arguments, so you don’t need to define the residuals any more.

Going further
• Try with a more complex waveform (for instance waveform_2.npy) that contains three significant
peaks. You must adapt the model which is now a sum of Gaussian functions instead of only one
Gaussian peak.
• In some cases, writing an explicit function to compute the Jacobian is faster than letting leastsq
estimate it numerically. Create a function to compute the Jacobian of the residuals and use it as

5.11. Summary exercises on scientific computing 244 5.11. Summary exercises on scientific computing 245
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

5.11.3 Image processing application: counting bubbles and unmolten grains 7. Attribute labels to all bubbles and sand grains, and remove from the sand mask grains that are
smaller than 10 pixels. To do so, use ndimage.sum or np.bincount to compute the grain sizes.
8. Compute the mean size of bubbles.

5.11.4 Example of solution for the image processing exercise: unmolten grains in
glass

Statement of the problem

1. Open the image file MV_HFV_012.jpg and display it. Browse through the keyword arguments
in the docstring of imshow to display the image with the “right” orientation (origin in the bottom
left corner, and not the upper left corner as for standard arrays).
This Scanning Element Microscopy image shows a glass sample (light gray matrix) with some
bubbles (on black) and unmolten sand grains (dark gray). We wish to determine the fraction
of the sample covered by these three phases, and to estimate the typical size of sand grains and 1. Open the image file MV_HFV_012.jpg and display it. Browse through the keyword arguments
bubbles, their sizes, etc. in the docstring of imshow to display the image with the “right” orientation (origin in the bottom
left corner, and not the upper left corner as for standard arrays).
2. Crop the image to remove the lower panel with measure information.
>>> dat = plt.imread('data/MV_HFV_012.jpg')
3. Slightly filter the image with a median filter in order to refine its histogram. Check how the
histogram changes.
2. Crop the image to remove the lower panel with measure information.
4. Using the histogram of the filtered image, determine thresholds that allow to define masks for sand
pixels, glass pixels and bubble pixels. Other option (homework): write a function that determines >>> dat = dat[:-60]
automatically the thresholds from the minima of the histogram.
3. Slightly filter the image with a median filter in order to refine its histogram. Check how the
5. Display an image in which the three phases are colored with three different colors.
histogram changes.
6. Use mathematical morphology to clean the different phases.

5.11. Summary exercises on scientific computing 246 5.11. Summary exercises on scientific computing 247
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

>>> filtdat = sp.ndimage.median_filter(dat, size=(7,7))

>>> hi_dat = np.histogram(dat, bins=np.arange(256))
>>> hi_filtdat = np.histogram(filtdat, bins=np.arange(256))

6. Use mathematical morphology to clean the different phases.

>>> sand_op = sp.ndimage.binary_opening(sand, iterations=2)

7. Attribute labels to all bubbles and sand grains, and remove from the sand mask grains that are
4. Using the histogram of the filtered image, determine thresholds that allow to define masks for sand smaller than 10 pixels. To do so, use sp.ndimage.sum or np.bincount to compute the grain sizes.
pixels, glass pixels and bubble pixels. Other option (homework): write a function that determines
automatically the thresholds from the minima of the histogram. >>> sand_labels, sand_nb = sp.ndimage.label(sand_op)
>>> sand_areas = np.array(sp.ndimage.sum(sand_op, sand_labels, np.arange(sand_
>>> void = filtdat <= 50 ˓→labels.max()+1)))
>>> sand = np.logical_and(filtdat > 50, filtdat <= 114) >>> mask = sand_areas > 100
>>> glass = filtdat > 114 >>> remove_small_sand = mask[sand_labels.ravel()].reshape(sand_labels.shape)

5. Display an image in which the three phases are colored with three different colors.

>>> phases = void.astype(int) + 2*glass.astype(int) + 3*sand.astype(int)

8. Compute the mean size of bubbles.

>>> bubbles_labels, bubbles_nb = sp.ndimage.label(void)

>>> bubbles_areas = np.bincount(bubbles_labels.ravel())[1:]
>>> mean_bubble_size = bubbles_areas.mean()
>>> median_bubble_size = np.median(bubbles_areas)
(continues on next page)

5.11. Summary exercises on scientific computing 248 5.11. Summary exercises on scientific computing 249
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

import scipy as sp
>>> mean_bubble_size, median_bubble_size
(np.float64(1699.875), np.float64(65.0)) # The default (Nelder Mead)
print(sp.optimize.minimize(f, x0=0))

message: Optimization terminated successfully.

5.12 Full code examples for the SciPy chapter success: True
status: 0
5.12.1 Finding the minimum of a smooth function fun: -7.945823375615215
x: [-1.306e+00]
Demos various methods to find the minimum of a function. nit: 5
jac: [-1.192e-06]
import numpy as np hess_inv: [[ 8.589e-02]]
import matplotlib.pyplot as plt nfev: 12
njev: 6

def f(x):
return x**2 + 10 * np.sin(x) plt.show()

Total running time of the script: (0 minutes 0.047 seconds)

x = np.arange(-5, 5, 0.1)
plt.plot(x, f(x))
5.12.2 Resample a signal with scipy.signal.resample
scipy.signal.resample() uses FFT to resample a 1D signal.
Generate a signal with 100 data point

import numpy as np

t = np.linspace(0, 5, 100)
x = np.sin(t)

Downsample it by a factor of 4

import scipy as sp

x_resampled = sp.signal.resample(x, 25)

Plot

import matplotlib.pyplot as plt

plt.figure(figsize=(5, 4))
plt.plot(t, x, label="Original signal")
plt.plot(t[::4], x_resampled, "ko", label="Resampled signal")

plt.legend(loc="best")
plt.show()

[<matplotlib.lines.Line2D object at 0x7f24887a4bc0>]

Now find the minimum with a few methods

5.12. Full code examples for the SciPy chapter 250 5.12. Full code examples for the SciPy chapter 251
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Total running time of the script: (0 minutes 0.249 seconds) Total running time of the script: (0 minutes 0.052 seconds)

5.12.3 Detrending a signal 5.12.4 Integrating a simple ODE

scipy.signal.detrend() removes a linear trend. Solve the ODE dy/dt = -2y between t = 0..4, with the initial condition y(t=0) = 1.
Generate a random signal with a trend

import numpy as np

t = np.linspace(0, 5, 100)
rng = np.random.default_rng()
x = t + rng.normal(size=100)

Detrend

import scipy as sp

x_detrended = sp.signal.detrend(x)

Plot

import matplotlib.pyplot as plt

plt.figure(figsize=(5, 4))
plt.plot(t, x, label="x")
plt.plot(t, x_detrended, label="x_detrended") import numpy as np
plt.legend(loc="best") import scipy as sp
plt.show() import matplotlib.pyplot as plt

def f(t, y):

return -2 * y
(continues on next page)

5.12. Full code examples for the SciPy chapter 252 5.12. Full code examples for the SciPy chapter 253
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

import matplotlib.pyplot as plt

t_span = (0, 4) # time interval dist = sp.stats.norm(loc=0, scale=1) # standard normal distribution
t_eval = np.linspace(*t_span) # times at which to evaluate `y` sample = dist.rvs(size=100000) # "random variate sample"
y0 = [ plt.hist(
1, sample,
] # initial state bins=51, # group the observations into 50 bins
res = sp.integrate.solve_ivp(f, t_span=t_span, y0=y0, t_eval=t_eval) density=True, # normalize the frequencies
label="normalized histogram",
plt.figure(figsize=(4, 3)) )
plt.plot(res.t, res.y[0])
plt.xlabel("t") x = np.linspace(-5, 5) # possible values of the random variable
plt.ylabel("y") plt.plot(x, dist.pdf(x), label="PDF")
plt.title("Solution of Initial Value Problem") plt.legend()
plt.tight_layout() plt.show()
plt.show()
Total running time of the script: (0 minutes 0.098 seconds)
Total running time of the script: (0 minutes 0.077 seconds)

5.12.6 Integrate the Damped spring-mass oscillator

5.12.5 Normal distribution: histogram and PDF
Explore the normal distribution: a histogram built from samples and the PDF (probability density
function).

import numpy as np
import scipy as sp
import matplotlib.pyplot as plt

m = 0.5 # kg
k = 4 # N/m
c = 0.4 # N s/m

zeta = c / (2 * m * np.sqrt(k / m))

omega = np.sqrt(k / m)

def f(t, z, zeta, omega):

return (z[1], -zeta * omega * z[1] - omega**2 * z[0])
import numpy as np
import scipy as sp
(continues on next page)
(continues on next page)

5.12. Full code examples for the SciPy chapter 254 5.12. Full code examples for the SciPy chapter 255
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

histogram2, bins = np.histogram(samples2, bins=bins, density=True)
t_span = (0, 10)
t_eval = np.linspace(*t_span, 100) plt.figure(figsize=(6, 4))
z0 = [1, 0] plt.hist(samples1, bins=bins, density=True, label="Samples 1") # type: ignore[arg-
res = sp.integrate.solve_ivp( ˓→type]

f, t_span, z0, t_eval=t_eval, args=(zeta, omega), method="LSODA" plt.hist(samples2, bins=bins, density=True, label="Samples 2") # type: ignore[arg-
) ˓→type]

plt.legend(loc="best")
plt.figure(figsize=(4, 3)) plt.show()
plt.plot(res.t, res.y[0], label="y")
plt.plot(res.t, res.y[1], label="dy/dt") Total running time of the script: (0 minutes 0.093 seconds)
plt.legend(loc="best")
plt.show()

Total running time of the script: (0 minutes 0.050 seconds) 5.12.8 Curve fitting
Demos a simple curve fitting
First generate some data
5.12.7 Comparing 2 sets of samples from Gaussians
import numpy as np

# Seed the random number generator for reproducibility

rng = np.random.default_rng(27446968)

x_data = np.linspace(-5, 5, num=50)

noise = 0.01 * np.cos(100 * x_data)
a, b = 2.9, 1.5
y_data = a * np.cos(b * x_data) + noise

# And plot it
import matplotlib.pyplot as plt

plt.figure(figsize=(6, 4))
plt.scatter(x_data, y_data)

import numpy as np
import matplotlib.pyplot as plt

# Generates 2 sets of observations

rng = np.random.default_rng(27446968)
samples1 = rng.normal(0, size=1000)
samples2 = rng.normal(1, size=1000)

# Compute a histogram of the sample

bins = np.linspace(-4, 4, 30)
histogram1, bins = np.histogram(samples1, bins=bins, density=True)
(continues on next page)

5.12. Full code examples for the SciPy chapter 256 5.12. Full code examples for the SciPy chapter 257
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

<matplotlib.collections.PathCollection object at 0x7f24a22ca2d0> Total running time of the script: (0 minutes 0.101 seconds)

Now fit a simple sine function to the data

import scipy as sp 5.12.9 Spectrogram, power spectral density

Demo spectrogram and power spectral density on a frequency chirp.

def test_func(x, a, b, c): import numpy as np

return a * np.sin(b * x + c) import matplotlib.pyplot as plt

params, params_covariance = sp.optimize.curve_fit( Generate a chirp signal

test_func, x_data, y_data, p0=[2, 1, 3]
# Seed the random number generator
)
np.random.seed(0)
print(params)
time_step = 0.01
time_vec = np.arange(0, 70, time_step)
[2.900026 1.50012043 1.57079633]
# A signal with a small frequency chirp
And plot the resulting curve on the data sig = np.sin(0.5 * np.pi * time_vec * (1 + 0.1 * time_vec))
plt.figure(figsize=(6, 4))
plt.figure(figsize=(8, 5))
plt.scatter(x_data, y_data, label="Data")
plt.plot(time_vec, sig)
plt.plot(x_data, test_func(x_data, *params), label="Fitted function")

plt.legend(loc="best")

plt.show()

5.12. Full code examples for the SciPy chapter 258 5.12. Full code examples for the SciPy chapter 259
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

[<matplotlib.lines.Line2D object at 0x7f24a23f9e50>]

Compute and plot the power spectral density (PSD)

Compute and plot the spectrogram The power of the signal per frequency band

The spectrum of the signal on consecutive time windows freqs, psd = sp.signal.welch(sig)

import scipy as sp plt.figure(figsize=(5, 4))

plt.semilogx(freqs, psd)
freqs, times, spectrogram = sp.signal.spectrogram(sig) plt.title("PSD: power spectral density")
plt.xlabel("Frequency")
plt.figure(figsize=(5, 4)) plt.ylabel("Power")
plt.imshow(spectrogram, aspect="auto", cmap="hot_r", origin="lower") plt.tight_layout()
plt.title("Spectrogram")
plt.ylabel("Frequency band")
plt.xlabel("Time window")
plt.tight_layout()

5.12. Full code examples for the SciPy chapter 260 5.12. Full code examples for the SciPy chapter 261
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

opened_mask = sp.ndimage.binary_opening(mask)
closed_mask = sp.ndimage.binary_closing(opened_mask)

# Plot
import matplotlib.pyplot as plt

plt.figure(figsize=(12, 3.5))
plt.subplot(141)
plt.imshow(a, cmap="gray")
plt.axis("off")
plt.title("a")

plt.subplot(142)
plt.imshow(mask, cmap="gray")
plt.axis("off")
plt.title("mask")

plt.subplot(143)
plt.imshow(opened_mask, cmap="gray")
plt.axis("off")
plt.title("opened_mask")

plt.subplot(144)
plt.show() plt.imshow(closed_mask, cmap="gray")
plt.title("closed_mask")
Total running time of the script: (0 minutes 0.327 seconds) plt.axis("off")

plt.subplots_adjust(wspace=0.05, left=0.01, bottom=0.01, right=0.99, top=0.99)

5.12.10 Demo mathematical morphology plt.show()

A basic demo of binary opening and closing.
Total running time of the script: (0 minutes 0.085 seconds)

5.12.11 Plot geometrical transformations on images

Demo geometrical transformations of images.

# Generate some binary data

import numpy as np

np.random.seed(0) Downloading file 'face.dat' from 'https://raw.githubusercontent.com/scipy/dataset-

˓→face/main/face.dat' to '/home/runner/.cache/scipy-data'.
a = np.zeros((50, 50))
a[10:-10, 10:-10] = 1
a += 0.25 * np.random.standard_normal(a.shape)
mask = a >= 0.5

# Apply mathematical morphology

import scipy as sp
(continues on next page)

5.12. Full code examples for the SciPy chapter 262 5.12. Full code examples for the SciPy chapter 263
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

# Load some data
import scipy as sp )
mask = sig > 1
face = sp.datasets.face(gray=True)
plt.figure(figsize=(7, 3.5))
# Apply a variety of transformations plt.subplot(1, 2, 1)
import matplotlib.pyplot as plt plt.imshow(sig)
plt.axis("off")
shifted_face = sp.ndimage.shift(face, (50, 50)) plt.title("sig")
shifted_face2 = sp.ndimage.shift(face, (50, 50), mode="nearest")
rotated_face = sp.ndimage.rotate(face, 30) plt.subplot(1, 2, 2)
cropped_face = face[50:-50, 50:-50] plt.imshow(mask, cmap="gray")
zoomed_face = sp.ndimage.zoom(face, 2) plt.axis("off")
zoomed_face.shape plt.title("mask")
plt.subplots_adjust(wspace=0.05, left=0.01, bottom=0.01, right=0.99, top=0.9)
plt.figure(figsize=(15, 3))
plt.subplot(151)
plt.imshow(shifted_face, cmap="gray")
plt.axis("off")

plt.subplot(152)
plt.imshow(shifted_face2, cmap="gray")
plt.axis("off")

plt.subplot(153)
plt.imshow(rotated_face, cmap="gray")
plt.axis("off")

plt.subplot(154)
plt.imshow(cropped_face, cmap="gray")
plt.axis("off")

plt.subplot(155)
plt.imshow(zoomed_face, cmap="gray")
plt.axis("off")
Label connected components
plt.subplots_adjust(wspace=0.05, left=0.01, bottom=0.01, right=0.99, top=0.99)
import scipy as sp
plt.show()
labels, nb = sp.ndimage.label(mask)
Total running time of the script: (0 minutes 1.012 seconds)
plt.figure(figsize=(3.5, 3.5))
plt.imshow(labels)
plt.title("label")
5.12.12 Demo connected components plt.axis("off")
Extracting and labeling connected components in a 2D array
plt.subplots_adjust(wspace=0.05, left=0.01, bottom=0.01, right=0.99, top=0.9)
import numpy as np
import matplotlib.pyplot as plt

Generate some binary data

x, y = np.indices((100, 100))
sig = (
np.sin(2 * np.pi * x / 50.0)
* np.sin(2 * np.pi * y / 50.0)
* (1 + x * y / 50.0**2) ** 2
(continues on next page)

5.12. Full code examples for the SciPy chapter 264 5.12. Full code examples for the SciPy chapter 265
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

5.12.13 Minima and roots of a function

Demos finding minima and roots of a function.

Define the function

import numpy as np

x = np.arange(-10, 10, 0.1)

def f(x):
return x**2 + 10 * np.sin(x)

Find minima

import scipy as sp

# Global optimization
grid = (-10, 10, 0.1)
xmin_global = sp.optimize.brute(f, (grid,))
print(f"Global minima found { xmin_global} ")
Extract the 4th connected component, and crop the array around it
# Constrain optimization
sl = sp.ndimage.find_objects(labels == 4) xmin_local = sp.optimize.fminbound(f, 0, 10)
plt.figure(figsize=(3.5, 3.5)) print(f"Local minimum found { xmin_local} ")
plt.imshow(sig[sl[0]])
plt.title("Cropped connected component")
plt.axis("off") Global minima found [-1.30641113]
Local minimum found 3.8374671194983834
plt.subplots_adjust(wspace=0.05, left=0.01, bottom=0.01, right=0.99, top=0.9)
Root finding
plt.show()
root = sp.optimize.root(f, 1) # our initial guess is 1
print(f"First root found { root.x} ")
root2 = sp.optimize.root(f, -2.5)
print(f"Second root found { root2.x} ")

First root found [0.]

Second root found [-2.47948183]

Plot function, minima, and roots

import matplotlib.pyplot as plt

fig = plt.figure(figsize=(6, 4))

ax = fig.add_subplot(111)

# Plot the function

ax.plot(x, f(x), "b-", label="f(x)")

# Plot the minima

xmins = np.array([xmin_global[0], xmin_local])
ax.plot(xmins, f(xmins), "go", label="Minima")

Total running time of the script: (0 minutes 0.100 seconds) # Plot the roots
roots = np.array([root.x, root2.x])
(continues on next page)

5.12. Full code examples for the SciPy chapter 266 5.12. Full code examples for the SciPy chapter 267
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

ax.plot(roots, f(roots), "kv", label="Roots") face = sp.datasets.face(gray=True)
face = face[:512, -512:] # crop out square on right
# Decorate the figure
ax.legend(loc="best") # Apply a variety of filters
ax.set_xlabel("x") import matplotlib.pyplot as plt
ax.set_ylabel("f(x)")
ax.axhline(0, color="gray") import numpy as np
plt.show()
noisy_face = np.copy(face).astype(float)
rng = np.random.default_rng()
noisy_face += face.std() * 0.5 * rng.standard_normal(face.shape)
blurred_face = sp.ndimage.gaussian_filter(noisy_face, sigma=3)
median_face = sp.ndimage.median_filter(noisy_face, size=5)
wiener_face = sp.signal.wiener(noisy_face, (5, 5))

plt.figure(figsize=(12, 3.5))
plt.subplot(141)
plt.imshow(noisy_face, cmap="gray")
plt.axis("off")
plt.title("noisy")

plt.subplot(142)
plt.imshow(blurred_face, cmap="gray")
plt.axis("off")
plt.title("Gaussian filter")

plt.subplot(143)
plt.imshow(median_face, cmap="gray")
plt.axis("off")
plt.title("median filter")

plt.subplot(144)
plt.imshow(wiener_face, cmap="gray")
plt.title("Wiener filter")
Total running time of the script: (0 minutes 0.064 seconds) plt.axis("off")

plt.subplots_adjust(wspace=0.05, left=0.01, bottom=0.01, right=0.99, top=0.99)

5.12.14 Plot filtering on images plt.show()
Demo filtering for denoising of images.
Total running time of the script: (0 minutes 0.492 seconds)

5.12.15 Optimization of a two-parameter function

import numpy as np

# Define the function that we are interested in

def sixhump(x):
return (
(4 - 2.1 * x[0] ** 2 + x[0] ** 4 / 3) * x[0] ** 2
# Load some data + x[0] * x[1]
import scipy as sp + (-4 + 4 * x[1] ** 2) * x[1] ** 2
)
(continues on next page) (continues on next page)

5.12. Full code examples for the SciPy chapter 268 5.12. Full code examples for the SciPy chapter 269
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

fig = plt.figure()
ax: Axes3D = fig.add_subplot(111, projection="3d")
# Make a grid to evaluate the function (for plotting) surf = ax.plot_surface(
xlim = [-2, 2] xg,
ylim = [-1, 1] yg,
x = np.linspace(*xlim) # type: ignore[call-overload] sixhump([xg, yg]),
y = np.linspace(*ylim) # type: ignore[call-overload] rstride=1,
xg, yg = np.meshgrid(x, y) cstride=1,
cmap="viridis",
linewidth=0,
A 2D image plot of the function antialiased=False,
Simple visualization in 2D )

import matplotlib.pyplot as plt ax.set_xlabel("x")

ax.set_ylabel("y")
plt.figure() ax.set_zlabel("f(x, y)")
plt.imshow(sixhump([xg, yg]), extent=xlim + ylim, origin="lower") # type: ignore[arg- ax.set_title("Six-hump Camelback function")
˓→type]

plt.colorbar()

Text(0.5, 1.0, 'Six-hump Camelback function')

<matplotlib.colorbar.Colorbar object at 0x7f24a2242480>

Find minima

A 3D surface plot of the function import scipy as sp

from mpl_toolkits.mplot3d import Axes3D # local minimization

res_local = sp.optimize.minimize(sixhump, x0=[0, 0])
(continues on next page) (continues on next page)

5.12. Full code examples for the SciPy chapter 270 5.12. Full code examples for the SciPy chapter 271
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) Generate the signal

# Seed the random number generator

# global minimization
rng = np.random.default_rng(27446968)
res_global = sp.optimize.differential_evolution(sixhump, bounds=[xlim, ylim])
time_step = 0.02
plt.figure()
period = 5.0
# Show the function in 2D
plt.imshow(sixhump([xg, yg]), extent=xlim + ylim, origin="lower") # type: ignore[arg-
time_vec = np.arange(0, 20, time_step)
˓→type]
sig = np.sin(2 * np.pi / period * time_vec) + 0.5 * rng.normal(size=time_vec.size)
plt.colorbar()
# Mark the minima
plt.figure(figsize=(6, 5))
plt.scatter(res_local.x[0], res_local.x[1], label="local minimizer")
plt.plot(time_vec, sig, label="Original signal")
plt.scatter(res_global.x[0], res_global.x[1], label="global minimizer")
plt.legend()
plt.show()

Total running time of the script: (0 minutes 0.393 seconds) [<matplotlib.lines.Line2D object at 0x7f24a2a07110>]

Compute and plot the power

5.12.16 Plotting and manipulating FFTs for filtering # The FFT of the signal
Plot the power of the FFT of a signal and inverse FFT back to reconstruct a signal. sig_fft = sp.fft.fft(sig)

This example demonstrate scipy.fft.fft(), scipy.fft.fftfreq() and scipy.fft.ifft(). It imple- # And the power (sig_fft is of complex dtype)
ments a basic filter that is very suboptimal, and should not be used. power = np.abs(sig_fft) ** 2
import numpy as np
# The corresponding frequencies
import scipy as sp
sample_freq = sp.fft.fftfreq(sig.size, d=time_step)
import matplotlib.pyplot as plt
(continues on next page)

5.12. Full code examples for the SciPy chapter 272 5.12. Full code examples for the SciPy chapter 273
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

[]

# Plot the FFT power

plt.figure(figsize=(6, 5)) Remove all the high frequencies
plt.plot(sample_freq, power)
We now remove all the high frequencies and transform back from frequencies to signal.
plt.xlabel("Frequency [Hz]")
plt.ylabel("plower") high_freq_fft = sig_fft.copy()
high_freq_fft[np.abs(sample_freq) > peak_freq] = 0
# Find the peak frequency: we can focus on only the positive frequencies filtered_sig = sp.fft.ifft(high_freq_fft)
pos_mask = np.where(sample_freq > 0)
freqs = sample_freq[pos_mask] plt.figure(figsize=(6, 5))
peak_freq = freqs[power[pos_mask].argmax()] plt.plot(time_vec, sig, label="Original signal")
plt.plot(time_vec, filtered_sig, linewidth=3, label="Filtered signal")
# Check that it does indeed correspond to the frequency that we generate plt.xlabel("Time [s]")
# the signal with plt.ylabel("Amplitude")
np.allclose(peak_freq, 1.0 / period)
plt.legend(loc="best")
# An inner plot to show the peak frequency
axes = plt.axes((0.55, 0.3, 0.3, 0.5))
plt.title("Peak frequency")
plt.plot(freqs[:8], power[pos_mask][:8])
plt.setp(axes, yticks=[])

# scipy.signal.find_peaks_cwt can also be used for more advanced

# peak detection

/opt/hostedtoolcache/Python/3.12.9/x64/lib/python3.12/site-packages/matplotlib/cbook.
˓→py:1719: ComplexWarning: Casting complex values to real discards the imaginary part

return math.isfinite(val)
/opt/hostedtoolcache/Python/3.12.9/x64/lib/python3.12/site-packages/matplotlib/cbook.
˓→py:1355: ComplexWarning: Casting complex values to real discards the imaginary part

return np.asarray(x, float)

<matplotlib.legend.Legend object at 0x7f24a2309340>

5.12. Full code examples for the SciPy chapter 274 5.12. Full code examples for the SciPy chapter 275
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Note This is actually a bad way of creating a filter: such brutal cut-off in frequency space does not
control distortion on the signal.
Filters should be created using the SciPy filter design code

plt.show()

Total running time of the script: (0 minutes 0.186 seconds)

5.12.17 A demo of 1D interpolation

# Generate data
import numpy as np

rng = np.random.default_rng(27446968)
measured_time = np.linspace(0, 2 * np.pi, 20)
function = np.sin(measured_time)
noise = rng.normal(loc=0, scale=0.1, size=20)
measurements = function + noise
(continues on next page)

5.12. Full code examples for the SciPy chapter 276 5.12. Full code examples for the SciPy chapter 277
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) Load the data

# Smooth the curve and interpolate at new times import numpy as np

import scipy as sp
data = np.loadtxt("../../../../data/populations.txt")
smoothing_spline = sp.interpolate.make_smoothing_spline(measured_time, measurements) years = data[:, 0]
interpolation_time = np.linspace(0, 2 * np.pi, 200) populations = data[:, 1:]
smooth_results = smoothing_spline(interpolation_time)
Plot the data
# Plot the data, the interpolant, and the original function
import matplotlib.pyplot as plt
import matplotlib.pyplot as plt
plt.figure(figsize=(6, 4))
plt.figure()
plt.plot(measured_time, measurements, ".", ms=6, label="measurements")
plt.plot(years, populations * 1e-3)
plt.plot(interpolation_time, smooth_results, label="smoothing spline")
plt.xlabel("Year")
plt.plot(interpolation_time, np.sin(interpolation_time), "--", label="underlying curve
plt.ylabel(r"Population number ($\cdot10^3$)")
˓→")
plt.legend(["hare", "lynx", "carrot"], loc=1)
plt.legend()
plt.show()

# Fit the data exactly

interp_spline = sp.interpolate.make_interp_spline(measured_time, function)
interp_results = interp_spline(interpolation_time)

# Plot the data, the interpolant, and the original function

plt.figure(figsize=(6, 4))
plt.plot(measured_time, function, ".", ms=6, label="measurements")
plt.plot(interpolation_time, interp_results, label="interpolating spline")
plt.plot(interpolation_time, np.sin(interpolation_time), "--", label="underlying curve
˓→")

plt.legend()
plt.show()

# Plot interpolant, its derivative, and its antiderivative

plt.figure(figsize=(6, 4))
t = interpolation_time
plt.plot(t, interp_spline(t), label="spline")
plt.plot(t, interp_spline.derivative()(t), label="derivative")
plt.plot(t, interp_spline.antiderivative()(t) - 1, label="antiderivative")

plt.legend()
plt.show()

Total running time of the script: (0 minutes 0.186 seconds)

5.12.18 Solutions of the exercises for SciPy

<matplotlib.legend.Legend object at 0x7f24a2ce1f10>
Solutions of the exercises for SciPy
Crude periodicity finding Plot its periods
Discover the periods in evolution of animal populations (../../../../data/populations.txt) import scipy as sp

ft_populations = sp.fft.fft(populations, axis=0)

frequencies = sp.fft.fftfreq(populations.shape[0], years[1] - years[0])
periods = 1 / frequencies
(continues on next page)

5.12. Full code examples for the SciPy chapter 278 5.12. Full code examples for the SciPy chapter 279
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) The data

plt.figure() import numpy as np

plt.plot(periods, abs(ft_populations) * 1e-3, "o")
plt.xlim(0, 22) temp_max = np.array([17, 19, 21, 28, 33, 38, 37, 37, 31, 23, 19, 18])
plt.xlabel("Period") temp_min = np.array([-62, -59, -56, -46, -32, -18, -9, -13, -25, -46, -52, -58])
plt.ylabel(r"Power ($\cdot10^3$)")
import matplotlib.pyplot as plt
plt.show()
months = np.arange(12)
plt.plot(months, temp_max, "ro")
plt.plot(months, temp_min, "bo")
plt.xlabel("Month")
plt.ylabel("Min and max temperature")

/home/runner/work/scientific-python-lectures/scientific-python-lectures/intro/scipy/
˓→examples/solutions/plot_periodicity_finder.py:39: RuntimeWarning: divide by zero␣

˓→encountered in divide

periods = 1 / frequencies

There’s probably a period of around 10 years (obvious from the plot), but for this crude a method, Text(35.472222222222214, 0.5, 'Min and max temperature')
there’s not enough data to say much more.
Total running time of the script: (0 minutes 0.124 seconds) Fitting it to a periodic function

import scipy as sp

Curve fitting: temperature as a function of month of the year

def yearly_temps(times, avg, ampl, time_offset):
We have the min and max temperatures in Alaska for each months of the year. We would like to find a
return avg + ampl * np.cos((times + time_offset) * 2 * np.pi / times.max())
function to describe this yearly evolution.
For this, we will fit a periodic function.
res_max, cov_max = sp.optimize.curve_fit(yearly_temps, months, temp_max, [20, 10, 0])
res_min, cov_min = sp.optimize.curve_fit(yearly_temps, months, temp_min, [-40, 20, 0])

5.12. Full code examples for the SciPy chapter 280 5.12. Full code examples for the SciPy chapter 281
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Plotting the fit The original image

days = np.linspace(0, 12, num=365) # read image

img = plt.imread("../../../../data/elephant.png")
plt.figure() plt.figure()
plt.plot(months, temp_max, "ro") plt.imshow(img)
plt.plot(days, yearly_temps(days, *res_max), "r-")
plt.plot(months, temp_min, "bo")
plt.plot(days, yearly_temps(days, *res_min), "b-")
plt.xlabel("Month")
plt.ylabel(r"Temperature ($^\circ$C)")

plt.show()

<matplotlib.image.AxesImage object at 0x7f24a3292e10>

Prepare an Gaussian convolution kernel

# First a 1-D Gaussian

t = np.linspace(-10, 10, 30)
Total running time of the script: (0 minutes 0.099 seconds) bump = np.exp(-0.1 * t**2)
bump /= np.trapezoid(bump) # normalize the integral to 1

# make a 2-D kernel out of it

Simple image blur by convolution with a Gaussian kernel kernel = bump[:, np.newaxis] * bump[np.newaxis, :]

Blur an an image (../../../../data/elephant.png) using a Gaussian kernel.

Implement convolution via FFT
Convolution is easy to perform with FFT: convolving two signals boils down to multiplying their FFTs
(and performing an inverse FFT).
# Padded fourier transform, with the same shape as the image
import numpy as np # We use :func:`scipy.fft.fft2` to have a 2D FFT
import scipy as sp kernel_ft = sp.fft.fft2(kernel, s=img.shape[:2], axes=(0, 1))
import matplotlib.pyplot as plt
# convolve
img_ft = sp.fft.fft2(img, axes=(0, 1))
(continues on next page)

5.12. Full code examples for the SciPy chapter 282 5.12. Full code examples for the SciPy chapter 283
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

# the 'newaxis' is to match to color direction plt.figure()
img2_ft = kernel_ft[:, :, np.newaxis] * img_ft plt.imshow(img3)
img2 = sp.fft.ifft2(img2_ft, axes=(0, 1)).real

# clip values to range

img2 = np.clip(img2, 0, 1)

# plot output
plt.figure()
plt.imshow(img2)

<matplotlib.image.AxesImage object at 0x7f24a2152150>

Note that we still have a decay to zero at the border of the image. Using scipy.ndimage.
gaussian_filter() would get rid of this artifact

plt.show()

Total running time of the script: (0 minutes 0.311 seconds)

<matplotlib.image.AxesImage object at 0x7f24a2308e90>

Further exercise (only if you are familiar with this stuff):

Image denoising by FFT
A “wrapped border” appears in the upper left and top edges of the image. This is because the padding
is not done correctly, and does not take the kernel size into account (so the convolution “flows out of Denoise an image (../../../../data/moonlanding.png) by implementing a blur with an FFT.
bounds of the image”). Try to remove this artifact. Implements, via FFT, the following convolution:
∫︁
A function to do it: scipy.signal.fftconvolve() 𝑓1 (𝑡) = 𝑑𝑡′ 𝐾(𝑡 − 𝑡′ )𝑓0 (𝑡′ )
The above exercise was only for didactic reasons: there exists a function in scipy that will do
this for us, and probably do a better job: scipy.signal.fftconvolve() 𝑓˜1 (𝜔) = 𝐾(𝜔)
˜ 𝑓˜0 (𝜔)

# mode='same' is there to enforce the same output shape as input arrays

# (ie avoid border effects)
img3 = sp.signal.fftconvolve(img, kernel[:, :, np.newaxis], mode="same")
(continues on next page)

5.12. Full code examples for the SciPy chapter 284 5.12. Full code examples for the SciPy chapter 285
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Read and plot the image (continued from previous page)

plt.colorbar()
import numpy as np
import matplotlib.pyplot as plt
plt.figure()
im = plt.imread("../../../../data/moonlanding.png").astype(float) plot_spectrum(im_fft)
plt.title("Fourier transform")
plt.figure()
plt.imshow(im, "gray")
plt.title("Original image")

Text(0.5, 1.0, 'Fourier transform')

Text(0.5, 1.0, 'Original image') Filter in FFT

# In the lines following, we'll make a copy of the original spectrum and
Compute the 2d FFT of the input image # truncate coefficients.

import scipy as sp # Define the fraction of coefficients (in each direction) we keep
keep_fraction = 0.1
im_fft = sp.fft.fft2(im)
# Call ff a copy of the original transform. NumPy arrays have a copy
# Show the results # method for this purpose.
im_fft2 = im_fft.copy()

def plot_spectrum(im_fft): # Set r and c to be the number of rows and columns of the array.
from matplotlib.colors import LogNorm r, c = im_fft2.shape

# A logarithmic colormap # Set to zero all rows with indices between r*keep_fraction and
plt.imshow(np.abs(im_fft), norm=LogNorm(vmin=5)) # r*(1-keep_fraction):
(continues on next page) (continues on next page)

5.12. Full code examples for the SciPy chapter 286 5.12. Full code examples for the SciPy chapter 287
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

im_fft2[int(r * keep_fraction) : int(r * (1 - keep_fraction))] = 0

# Similarly with the columns:

im_fft2[:, int(c * keep_fraction) : int(c * (1 - keep_fraction))] = 0

plt.figure()
plot_spectrum(im_fft2)
plt.title("Filtered Spectrum")

Text(0.5, 1.0, 'Reconstructed Image')

Easier and better: scipy.ndimage.gaussian_filter()

Implementing filtering directly with FFTs is tricky and time consuming. We can use the
Gaussian filter from scipy.ndimage

im_blur = sp.ndimage.gaussian_filter(im, 4)

plt.figure()
Text(0.5, 1.0, 'Filtered Spectrum') plt.imshow(im_blur, "gray")
plt.title("Blurred image")
Reconstruct the final image plt.show()

# Reconstruct the denoised image from the filtered spectrum, keep only the
# real part for display.
im_new = sp.fft.ifft2(im_fft2).real

plt.figure()
plt.imshow(im_new, "gray")
plt.title("Reconstructed Image")

5.12. Full code examples for the SciPy chapter 288 5.12. Full code examples for the SciPy chapter 289
 Scientific Python Lectures, Edition 2025.1rc0.dev0

CHAPTER 6
Getting help and finding documentation

Total running time of the script: (0 minutes 0.731 seconds) Author: Emmanuelle Gouillart
Rather than knowing all functions in NumPy and SciPy, it is important to find rapidly information
throughout the documentation and the available help. Here are some ways to get information:
• In Ipython, help function opens the docstring of the function. Only type the beginning of the
function’s name and use tab completion to display the matching functions.
ã See also
In [1]: help(np.van<TAB>
References to go further
• Some chapters of the advanced and the packages and applications parts of the SciPy lectures In [2]: help(np.vander)
Help on _ArrayFunctionDispatcher in module numpy:
• The SciPy cookbook
vander(x, N=None, increasing=False)
Generate a Vandermonde matrix.

The columns of the output matrix are powers of the input vector. The
order of the powers is determined by the `increasing` boolean argument.
Specifically, when `increasing` is False, the `i`-th output column is
the input vector raised element-wise to the power of ``N - i - 1``. Such
a matrix with a geometric progression in each row is named for Alexandre-
Theophile Vandermonde.

Parameters
----------
x : array_like
1-D input array.
N : int, optional
Number of columns in the output. If `N` is not specified, a square
array is returned (``N = len(x)``).
(continues on next page)

5.12. Full code examples for the SciPy chapter 290 291
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) packages.

increasing : bool, optional Tutorials on various topics as well as the complete API with all docstrings are found on this website.
Order of the powers of the columns. If True, the powers increase
from left to right, if False (the default) they are reversed. • Numpy’s and Scipy’s documentation is enriched and updated on a regular basis by users on a wiki
https://numpy.org/doc/stable/. As a result, some docstrings are clearer or more detailed on the
Returns wiki, and you may want to read directly the documentation on the wiki instead of the official
------- documentation website. Note that anyone can create an account on the wiki and write better
out : ndarray documentation; this is an easy way to contribute to an open-source project and improve the tools
Vandermonde matrix. If `increasing` is False, the first column is you are using!
``x^(N-1)``, the second ``x^(N-2)`` and so forth. If `increasing` is • The SciPy Cookbook https://scipy-cookbook.readthedocs.io gives recipes on many common prob-
True, the columns are ``x^0, x^1, ..., x^(N-1)``. lems frequently encountered, such as fitting data points, solving ODE, etc.

See Also • Matplotlib’s website https://matplotlib.org/ features a very nice gallery with a large number of
-------- plots, each of them shows both the source code and the resulting plot. This is very useful for
polynomial.polynomial.polyvander learning by example. More standard documentation is also available.
• In Ipython, the magical function %psearch search for objects matching patterns. This is useful if,
Examples for example, one does not know the exact name of a function.
--------
>>> import numpy as np In [3]: import numpy as np
>>> x = np.array([1, 2, 3, 5])
>>> N = 3 • If everything listed above fails (and Google doesn’t have the answer). . . don’t despair! There is
>>> np.vander(x, N) a vibrant Scientific Python community. Scientific Python is present on various platform. https:
array([[ 1, 1, 1], //scientific-python.org/community/
[ 4, 2, 1],
[ 9, 3, 1], Packages like SciPy and NumPy also have their own channels. Have a look at their respective
[25, 5, 1]]) websites to find out how to engage with users and maintainers.

>>> np.column_stack([x**(N-1-i) for i in range(N)])

array([[ 1, 1, 1],
[ 4, 2, 1],
[ 9, 3, 1],
[25, 5, 1]])

>>> x = np.array([1, 2, 3, 5])

>>> np.vander(x)
array([[ 1, 1, 1, 1],
[ 8, 4, 2, 1],
[ 27, 9, 3, 1],
[125, 25, 5, 1]])
>>> np.vander(x, increasing=True)
array([[ 1, 1, 1, 1],
[ 1, 2, 4, 8],
[ 1, 3, 9, 27],
[ 1, 5, 25, 125]])

The determinant of a square Vandermonde matrix is the product

of the differences between the values of the input vector:

>>> np.linalg.det(np.vander(x))
48.000000000000043 # may vary
>>> (5-3)*(5-2)*(5-1)*(3-2)*(3-1)*(2-1)
48

In Ipython it is not possible to open a separated window for help and documentation; however one can
always open a second Ipython shell just to display help and docstrings. . .
• Numpy’s and Scipy’s documentations can be browsed online on https://scipy.org and https://
numpy.org. The search button is quite useful inside the reference documentation of the two

292 293
 Scientific Python Lectures, Edition 2025.1rc0.dev0

This part of the Scientific Python Lectures is dedicated to advanced usage. It strives to educate the
proficient Python coder to be an expert and tackles various specific topics.

Part II

Advanced topics

294 295
 Scientific Python Lectures, Edition 2025.1rc0.dev0

– Replacing or tweaking the original object

– Decorators implemented as classes and as functions
– Copying the docstring and other attributes of the original function

7
– Examples in the standard library
– Deprecation of functions
– A while-loop removing decorator
– A plugin registration system

CHAPTER • Context managers

– Catching exceptions
– Using generators to define context managers

7.1 Iterators, generator expressions and generators

Advanced Python Constructs 7.1.1 Iterators

Simplicity

Duplication of effort is wasteful, and replacing the various home-grown approaches with a standard
feature usually ends up making things more readable, and interoperable as well.
Guido van Rossum — Adding Optional Static Typing to Python
Author Zbigniew Jędrzejewski-Szmek
An iterator is an object adhering to the iterator protocol — basically this means that it has a next
This section covers some features of the Python language which can be considered advanced — in method, which, when called, returns the next item in the sequence, and when there’s nothing to return,
the sense that not every language has them, and also in the sense that they are more useful in more raises the StopIteration exception.
complicated programs or libraries, but not in the sense of being particularly specialized, or particularly
complicated. An iterator object allows to loop just once. It holds the state (position) of a single iteration, or from the
other side, each loop over a sequence requires a single iterator object. This means that we can iterate
It is important to underline that this chapter is purely about the language itself — about features over the same sequence more than once concurrently. Separating the iteration logic from the sequence
supported through special syntax complemented by functionality of the Python stdlib, which could not allows us to have more than one way of iteration.
be implemented through clever external modules.
Calling the __iter__ method on a container to create an iterator object is the most straightforward way
The process of developing the Python programming language, its syntax, is very transparent; proposed to get hold of an iterator. The iter function does that for us, saving a few keystrokes.
changes are evaluated from various angles and discussed via Python Enhancement Proposals — PEPs.
As a result, features described in this chapter were added after it was shown that they indeed solve real >>> nums = [1, 2, 3] # note that ... varies: these are different objects
problems and that their use is as simple as possible. >>> iter(nums)
<...iterator object at ...>
>>> nums.__iter__()
Chapter contents <...iterator object at ...>
>>> nums.__reversed__()
• Iterators, generator expressions and generators <...reverseiterator object at ...>
– Iterators
>>> it = iter(nums)
– Generator expressions >>> next(it)
– Generators 1
>>> next(it)
– Bidirectional communication 2
– Chaining generators >>> next(it)
3
• Decorators >>> next(it)
Traceback (most recent call last):
(continues on next page)

296 7.1. Iterators, generator expressions and generators 297

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) A third way to create iterator objects is to call a generator function. A generator is a function containing
File "<stdin>", line 1, in <module> the keyword yield. It must be noted that the mere presence of this keyword completely changes the
StopIteration nature of the function: this yield statement doesn’t have to be invoked, or even reachable, but causes
the function to be marked as a generator. When a normal function is called, the instructions contained in
When used in a loop, StopIteration is swallowed and causes the loop to finish. But with explicit the body start to be executed. When a generator is called, the execution stops before the first instruction
invocation, we can see that once the iterator is exhausted, accessing it raises an exception. in the body. An invocation of a generator function creates a generator object, adhering to the iterator
protocol. As with normal function invocations, concurrent and recursive invocations are allowed.
Using the for..in loop also uses the __iter__ method. This allows us to transparently start the iteration
over a sequence. But if we already have the iterator, we want to be able to use it in an for loop in When next is called, the function is executed until the first yield. Each encountered yield statement
the same way. In order to achieve this, iterators in addition to next are also required to have a method gives a value becomes the return value of next. After executing the yield statement, the execution of
called __iter__ which returns the iterator (self). this function is suspended.

Support for iteration is pervasive in Python: all sequences and unordered containers in the standard >>> def f():
library allow this. The concept is also stretched to other things: e.g. file objects support iteration over ... yield 1
lines. ... yield 2
>>> f()
>>> with open("/etc/fstab") as f: <generator object f at 0x...>
... f is f.__iter__() >>> gen = f()
... >>> next(gen)
True 1
>>> next(gen)
The file is an iterator itself and it’s __iter__ method doesn’t create a separate object: only a single 2
thread of sequential access is allowed. >>> next(gen)
Traceback (most recent call last):
7.1.2 Generator expressions File "<stdin>", line 1, in <module>
StopIteration
A second way in which iterator objects are created is through generator expressions, the basis for list
comprehensions. To increase clarity, a generator expression must always be enclosed in parentheses Let’s go over the life of the single invocation of the generator function.
or an expression. If round parentheses are used, then a generator iterator is created. If rectangular
parentheses are used, the process is short-circuited and we get a list. >>> def f():
... print("-- start --")
>>> (i for i in nums) ... yield 3
<generator object <genexpr> at 0x...> ... print("-- finish --")
>>> [i for i in nums] ... yield 4
[1, 2, 3] >>> gen = f()
>>> list(i for i in nums) >>> next(gen)
[1, 2, 3] -- start --
3
The list comprehension syntax also extends to dictionary and set comprehensions. A set is cre- >>> next(gen)
ated when the generator expression is enclosed in curly braces. A dict is created when the generator -- finish --
expression contains “pairs” of the form key:value: 4
>>> next(gen)
>>> {i for i in range(3)}
Traceback (most recent call last):
{0, 1, 2}
...
>>> {i:i**2 for i in range(3)}
StopIteration
{0: 0, 1: 1, 2: 4}
Contrary to a normal function, where executing f() would immediately cause the first print to be
One gotcha should be mentioned: in old Pythons the index variable (i) would leak, and in versions >=
executed, gen is assigned without executing any statements in the function body. Only when gen.
3 this is fixed.
__next__() is invoked by next, the statements up to the first yield are executed. The second next
prints -- finish -- and execution halts on the second yield. The third next falls of the end of the
7.1.3 Generators function. Since no yield was reached, an exception is raised.
What happens with the function after a yield, when the control passes to the caller? The state of
Generators each generator is stored in the generator object. From the point of view of the generator function, is
looks almost as if it was running in a separate thread, but this is just an illusion: execution is strictly
A generator is a function that produces a sequence of results instead of a single value. single-threaded, but the interpreter keeps and restores the state in between the requests for the next
David Beazley — A Curious Course on Coroutines and Concurrency value.
Why are generators useful? As noted in the parts about iterators, a generator function is just a different
way to create an iterator object. Everything that can be done with yield statements, could also be

7.1. Iterators, generator expressions and generators 298 7.1. Iterators, generator expressions and generators 299
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

done with next methods. Nevertheless, using a function and having the interpreter perform its magic (continued from previous page)
to create an iterator has advantages. A function can be much shorter than the definition of a class with
the required next and __iter__ methods. What is more important, it is easier for the author of the >>> it = g()
generator to understand the state which is kept in local variables, as opposed to instance attributes, >>> next(it)
which have to be used to pass data between consecutive invocations of next on an iterator object. --start--
A broader question is why are iterators useful? When an iterator is used to power a loop, the loop --yielding 0--
becomes very simple. The code to initialise the state, to decide if the loop is finished, and to find the 0
next value is extracted into a separate place. This highlights the body of the loop — the interesting >>> it.send(11)
part. In addition, it is possible to reuse the iterator code in other places. --yield returned 11--
--yielding 1--
1
7.1.4 Bidirectional communication >>> it.throw(IndexError)
Each yield statement causes a value to be passed to the caller. This is the reason for the introduction --yield raised IndexError()--
of generators by PEP 255. But communication in the reverse direction is also useful. One obvious way --yielding 2--
would be some external state, either a global variable or a shared mutable object. Direct communication 2
is possible thanks to PEP 342. It is achieved by turning the previously boring yield statement into an >>> it.close()
expression. When the generator resumes execution after a yield statement, the caller can call a method --closing--
on the generator object to either pass a value into the generator, which then is returned by the yield
statement, or a different method to inject an exception into the generator.
7.1.5 Chaining generators
The first of the new methods is send(value), which is similar to next(), but passes value into the
generator to be used for the value of the yield expression. In fact, g.next() and g.send(None) are
equivalent. ò Note
The second of the new methods is throw(type, value=None, traceback=None) which is equivalent to: This is a preview of PEP 380 (not yet implemented, but accepted for Python 3.3).
raise type, value, traceback
Let’s say we are writing a generator and we want to yield a number of values generated by a second
at the point of the yield statement. generator, a subgenerator. If yielding of values is the only concern, this can be performed without
Unlike raise (which immediately raises an exception from the current execution point), throw() first much difficulty using a loop such as
resumes the generator, and only then raises the exception. The word throw was picked because it subgen = some_other_generator()
is suggestive of putting the exception in another location, and is associated with exceptions in other for v in subgen:
languages. yield v
What happens when an exception is raised inside the generator? It can be either raised explicitly or
when executing some statements or it can be injected at the point of a yield statement by means of However, if the subgenerator is to interact properly with the caller in the case of calls to send(), throw()
the throw() method. In either case, such an exception propagates in the standard manner: it can and close(), things become considerably more difficult. The yield statement has to be guarded by a
be intercepted by an except or finally clause, or otherwise it causes the execution of the generator try..except..finally structure similar to the one defined in the previous section to “debug” the generator
function to be aborted and propagates in the caller. function. Such code is provided in PEP 380#id13, here it suffices to say that new syntax to properly
yield from a subgenerator is being introduced in Python 3.3:
For completeness’ sake, it’s worth mentioning that generator iterators also have a close() method,
which can be used to force a generator that would otherwise be able to provide more values to finish yield from some_other_generator()
immediately. It allows the generator __del__ method to destroy objects holding the state of generator.
Let’s define a generator which just prints what is passed in through send and throw. This behaves like the explicit loop above, repeatedly yielding values from some_other_generator until
it is exhausted, but also forwards send, throw and close to the subgenerator.
>>> import itertools
>>> def g():
... print('--start--') 7.2 Decorators
... for i in itertools.count():
... print('--yielding %i --' % i)
... try: Summary
... ans = yield i
... except GeneratorExit: This amazing feature appeared in the language almost apologetically and with concern that it might
... print('--closing--') not be that useful.
... raise
Bruce Eckel — An Introduction to Python Decorators
... except Exception as e:
... print('--yield raised %r --' % e)
... else: Since functions and classes are objects, they can be passed around. Since they are mutable objects, they
... print('--yield returned %s --' % ans) can be modified. The act of altering a function or class object after it has been constructed but before
(continues on next page) is is bound to its name is called decorating.

7.1. Iterators, generator expressions and generators 300 7.2. Decorators 301
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

There are two things hiding behind the name “decorator” — one is the function which does the work of 7.2.2 Decorators implemented as classes and as functions
decorating, i.e. performs the real work, and the other one is the expression adhering to the decorator
The only requirement on decorators is that they can be called with a single argument. This means that
syntax, i.e. an at-symbol and the name of the decorating function.
decorators can be implemented as normal functions, or as classes with a __call__ method, or in theory,
Function can be decorated by using the decorator syntax for functions: even as lambda functions.

@decorator # ❷ Let’s compare the function and class approaches. The decorator expression (the part after @) can be
def function(): # ❶ either just a name, or a call. The bare-name approach is nice (less to type, looks cleaner, etc.), but is
pass only possible when no arguments are needed to customise the decorator. Decorators written as functions
can be used in those two cases:
• A function is defined in the standard way. ❶ >>> def simple_decorator(function):
• An expression starting with @ placed before the function definition is the decorator ❷. The part ... print("doing decoration")
after @ must be a simple expression, usually this is just the name of a function or class. This part is ... return function
evaluated first, and after the function defined below is ready, the decorator is called with the newly >>> @simple_decorator
defined function object as the single argument. The value returned by the decorator is attached to ... def function():
the original name of the function. ... print("inside function")
doing decoration
Decorators can be applied to functions and to classes. For classes the semantics are identical — the >>> function()
original class definition is used as an argument to call the decorator and whatever is returned is assigned inside function
under the original name.
Before the decorator syntax was implemented (PEP 318), it was possible to achieve the same effect by >>> def decorator_with_arguments(arg):
assigning the function or class object to a temporary variable and then invoking the decorator explicitly ... print("defining the decorator")
and then assigning the return value to the name of the function. This sounds like more typing, and it ... def _decorator(function):
is, and also the name of the decorated function doubling as a temporary variable must be used at least ... # in this inner function, arg is available too
three times, which is prone to errors. Nevertheless, the example above is equivalent to: ... print("doing decoration, %r " % arg)
... return function
def function(): # ❶ ... return _decorator
pass >>> @decorator_with_arguments("abc")
function = decorator(function) # ❷ ... def function():
... print("inside function")
Decorators can be stacked — the order of application is bottom-to-top, or inside-out. The semantics defining the decorator
are such that the originally defined function is used as an argument for the first decorator, whatever is doing decoration, 'abc'
returned by the first decorator is used as an argument for the second decorator, . . . , and whatever is >>> function()
returned by the last decorator is attached under the name of the original function. inside function
The decorator syntax was chosen for its readability. Since the decorator is specified before the header
of the function, it is obvious that its is not a part of the function body and its clear that it can only The two trivial decorators above fall into the category of decorators which return the original function.
operate on the whole function. Because the expression is prefixed with @ is stands out and is hard to If they were to return a new function, an extra level of nestedness would be required. In the worst case,
miss (“in your face”, according to the PEP :) ). When more than one decorator is applied, each one is three levels of nested functions.
placed on a separate line in an easy to read way. >>> def replacing_decorator_with_args(arg):
... print("defining the decorator")
7.2.1 Replacing or tweaking the original object ... def _decorator(function):
... # in this inner function, arg is available too
Decorators can either return the same function or class object or they can return a completely different
... print("doing decoration, %r " % arg)
object. In the first case, the decorator can exploit the fact that function and class objects are mutable
... def _wrapper(*args, **kwargs):
and add attributes, e.g. add a docstring to a class. A decorator might do something useful even without
... print("inside wrapper, %r %r " % (args, kwargs))
modifying the object, for example register the decorated class in a global registry. In the second case,
... return function(*args, **kwargs)
virtually anything is possible: when something different is substituted for the original function or class,
... return _wrapper
the new object can be completely different. Nevertheless, such behaviour is not the purpose of decorators:
... return _decorator
they are intended to tweak the decorated object, not do something unpredictable. Therefore, when a
>>> @replacing_decorator_with_args("abc")
function is “decorated” by replacing it with a different function, the new function usually calls the original
... def function(*args, **kwargs):
function, after doing some preparatory work. Likewise, when a class is “decorated” by replacing if with
... print("inside function, %r %r " % (args, kwargs))
a new class, the new class is usually derived from the original class. When the purpose of the decorator
... return 14
is to do something “every time”, like to log every call to a decorated function, only the second type of
defining the decorator
decorators can be used. On the other hand, if the first type is sufficient, it is better to use it, because it
doing decoration, 'abc'
is simpler.
>>> function(11, 12)
inside wrapper, (11, 12) {}
inside function, (11, 12) {}
(continues on next page)

7.2. Decorators 302 7.2. Decorators 303

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

14 ... print("in function, %s %s " % (args, kwargs))
in decorator call, foo
The _wrapper function is defined to accept all positional and keyword arguments. In general we cannot >>> function(11, 12)
know what arguments the decorated function is supposed to accept, so the wrapper function just passes in the wrapper, (11, 12) {}
everything to the wrapped function. One unfortunate consequence is that the apparent argument list is in function, (11, 12) {}
misleading.
A decorator like this can do pretty much anything, since it can modify the original function object and
Compared to decorators defined as functions, complex decorators defined as classes are simpler. When
mangle the arguments, call the original function or not, and afterwards mangle the return value.
an object is created, the __init__ method is only allowed to return None, and the type of the created
object cannot be changed. This means that when a decorator is defined as a class, it doesn’t make
much sense to use the argument-less form: the final decorated object would just be an instance of the 7.2.3 Copying the docstring and other attributes of the original function
decorating class, returned by the constructor call, which is not very useful. Therefore it’s enough to When a new function is returned by the decorator to replace the original function, an unfortunate
discuss class-based decorators where arguments are given in the decorator expression and the decorator consequence is that the original function name, the original docstring, the original argument list are
__init__ method is used for decorator construction. lost. Those attributes of the original function can partially be “transplanted” to the new function
>>> class decorator_class(object): by setting __doc__ (the docstring), __module__ and __name__ (the full name of the function), and
... def __init__(self, arg): __annotations__ (extra information about arguments and the return value of the function available in
... # this method is called in the decorator expression Python 3). This can be done automatically by using functools.update_wrapper.
... print("in decorator init, %s " % arg)
... self.arg = arg functools.update_wrapper(wrapper, wrapped)
... def __call__(self, function):
... # this method is called to do the job “Update a wrapper function to look like the wrapped function.”
... print("in decorator call, %s " % self.arg)
>>> import functools
... return function
>>> def replacing_decorator_with_args(arg):
>>> deco_instance = decorator_class('foo')
... print("defining the decorator")
in decorator init, foo
... def _decorator(function):
>>> @deco_instance
... print("doing decoration, %r " % arg)
... def function(*args, **kwargs):
... def _wrapper(*args, **kwargs):
... print("in function, %s %s " % (args, kwargs))
... print("inside wrapper, %r %r " % (args, kwargs))
in decorator call, foo
... return function(*args, **kwargs)
>>> function()
... return functools.update_wrapper(_wrapper, function)
in function, () {}
... return _decorator
>>> @replacing_decorator_with_args("abc")
Contrary to normal rules (PEP 8) decorators written as classes behave more like functions and therefore ... def function():
their name often starts with a lowercase letter. ... "extensive documentation"
In reality, it doesn’t make much sense to create a new class just to have a decorator which returns the ... print("inside function")
original function. Objects are supposed to hold state, and such decorators are more useful when the ... return 14
decorator returns a new object. defining the decorator
doing decoration, 'abc'
>>> class replacing_decorator_class(object): >>> function
... def __init__(self, arg): <function function at 0x...>
... # this method is called in the decorator expression >>> print(function.__doc__)
... print("in decorator init, %s " % arg) extensive documentation
... self.arg = arg
... def __call__(self, function):
... # this method is called to do the job One important thing is missing from the list of attributes which can be copied to the replacement
... print("in decorator call, %s " % self.arg) function: the argument list. The default values for arguments can be modified through the __defaults__,
... self.function = function __kwdefaults__ attributes, but unfortunately the argument list itself cannot be set as an attribute. This
... return self._wrapper means that help(function) will display a useless argument list which will be confusing for the user of
... def _wrapper(self, *args, **kwargs): the function. An effective but ugly way around this problem is to create the wrapper dynamically, using
... print("in the wrapper, %s %s " % (args, kwargs)) eval. This can be automated by using the external decorator module. It provides support for the
... return self.function(*args, **kwargs) decorator decorator, which takes a wrapper and turns it into a decorator which preserves the function
>>> deco_instance = replacing_decorator_class('foo') signature.
in decorator init, foo To sum things up, decorators should always use functools.update_wrapper or some other means of
>>> @deco_instance copying function attributes.
... def function(*args, **kwargs):
(continues on next page)

7.2. Decorators 304 7.2. Decorators 305

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

7.2.4 Examples in the standard library (continued from previous page)

First, it should be mentioned that there’s a number of useful decorators available in the standard library. def area(self, area):
There are three decorators which really form a part of the language: self.edge = area ** 0.5

• classmethod causes a method to become a “class method”, which means that it can be invoked The way that this works, is that the property decorator replaces the getter method with a property
without creating an instance of the class. When a normal method is invoked, the interpreter inserts object. This object in turn has three methods, getter, setter, and deleter, which can be used
the instance object as the first positional parameter, self. When a class method is invoked, the as decorators. Their job is to set the getter, setter and deleter of the property object (stored as
class itself is given as the first parameter, often called cls. attributes fget, fset, and fdel). The getter can be set like in the example above, when creating
Class methods are still accessible through the class’ namespace, so they don’t pollute the module’s the object. When defining the setter, we already have the property object under area, and we add
namespace. Class methods can be used to provide alternative constructors: the setter to it by using the setter method. All this happens when we are creating the class.
Afterwards, when an instance of the class has been created, the property object is special. When the
class Array(object):
interpreter executes attribute access, assignment, or deletion, the job is delegated to the methods
def __init__(self, data):
of the property object.
self.data = data
To make everything crystal clear, let’s define a “debug” example:
@classmethod
def fromfile(cls, file): >>> class D(object):
data = numpy.load(file) ... @property
return cls(data) ... def a(self):
... print("getting 1")
This is cleaner than using a multitude of flags to __init__. ... return 1
... @a.setter
• staticmethod is applied to methods to make them “static”, i.e. basically a normal function, but ... def a(self, value):
accessible through the class namespace. This can be useful when the function is only needed inside ... print("setting %r " % value)
this class (its name would then be prefixed with _), or when we want the user to think of the ... @a.deleter
method as connected to the class, despite an implementation which doesn’t require this. ... def a(self):
• property is the pythonic answer to the problem of getters and setters. A method decorated with ... print("deleting")
property becomes a getter which is automatically called on attribute access. >>> D.a
<property object at 0x...>
>>> class A(object): >>> D.a.fget
... @property <function ...>
... def a(self): >>> D.a.fset
... "an important attribute" <function ...>
... return "a value" >>> D.a.fdel
>>> A.a <function ...>
<property object at 0x...> >>> d = D() # ... varies, this is not the same `a` function
>>> A().a >>> d.a
'a value' getting 1
1
In this example, A.a is an read-only attribute. It is also documented: help(A) includes the >>> d.a = 2
docstring for attribute a taken from the getter method. Defining a as a property allows it to be a setting 2
calculated on the fly, and has the side effect of making it read-only, because no setter is defined. >>> del d.a
deleting
To have a setter and a getter, two methods are required, obviously: >>> d.a
class Rectangle(object): getting 1
def __init__(self, edge): 1
self.edge = edge
Properties are a bit of a stretch for the decorator syntax. One of the premises of the decorator
@property syntax — that the name is not duplicated — is violated, but nothing better has been invented so
def area(self): far. It is just good style to use the same name for the getter, setter, and deleter methods.
"""Computed area. Some newer examples include:

Setting this updates the edge length to the proper value. • functools.lru_cache memoizes an arbitrary function maintaining a limited cache of argu-
""" ments:answer pairs (Python 3.2)
return self.edge**2 • functools.total_ordering is a class decorator which fills in missing ordering methods (__lt__,
__gt__, __le__, . . . ) based on a single available one.
@area.setter
(continues on next page)

7.2. Decorators 306 7.2. Decorators 307

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

7.2.5 Deprecation of functions statements, but then the user would have to explicitly call list(find_answers()).
Let’s say we want to print a deprecation warning on stderr on the first invocation of a function we don’t We can define a decorator which constructs the list for us:
like anymore. If we don’t want to modify the function, we can use a decorator:
def vectorized(generator_func):
class deprecated(object): def wrapper(*args, **kwargs):
"""Print a deprecation warning once on first use of the function. return list(generator_func(*args, **kwargs))
return functools.update_wrapper(wrapper, generator_func)
>>> @deprecated() # doctest: +SKIP
... def f(): Our function then becomes:
... pass
>>> f() # doctest: +SKIP @vectorized
f is deprecated def find_answers():
""" while True:
def __call__(self, func): ans = look_for_next_answer()
self.func = func if ans is None:
self.count = 0 break
return self._wrapper yield ans
def _wrapper(self, *args, **kwargs):
self.count += 1
if self.count == 1: 7.2.7 A plugin registration system
print(self.func.__name__, 'is deprecated') This is a class decorator which doesn’t modify the class, but just puts it in a global registry. It falls into
return self.func(*args, **kwargs) the category of decorators returning the original object:

It can also be implemented as a function: class WordProcessor(object):

PLUGINS = []
def deprecated(func): def process(self, text):
"""Print a deprecation warning once on first use of the function. for plugin in self.PLUGINS:
text = plugin().cleanup(text)
>>> @deprecated # doctest: +SKIP return text
... def f():
... pass @classmethod
>>> f() # doctest: +SKIP def plugin(cls, plugin):
f is deprecated cls.PLUGINS.append(plugin)
"""
count = [0] @WordProcessor.plugin
def wrapper(*args, **kwargs): class CleanMdashesExtension(object):
count[0] += 1 def cleanup(self, text):
if count[0] == 1: return text.replace('&mdash;', u'\N{em dash}')
print(func.__name__, 'is deprecated')
return func(*args, **kwargs) Here we use a decorator to decentralise the registration of plugins. We call our decorator with a noun,
return wrapper instead of a verb, because we use it to declare that our class is a plugin for WordProcessor. Method
plugin simply appends the class to the list of plugins.

7.2.6 A while-loop removing decorator A word about the plugin itself: it replaces HTML entity for em-dash with a real Unicode em-dash
character. It exploits the unicode literal notation to insert a character by using its name in the unicode
Let’s say we have function which returns a lists of things, and this list created by running a loop. If we database (“EM DASH”). If the Unicode character was inserted directly, it would be impossible to
don’t know how many objects will be needed, the standard way to do this is something like: distinguish it from an en-dash in the source of a program.
def find_answers():
answers = [] ã See also
while True:
ans = look_for_next_answer() More examples and reading
if ans is None:
• PEP 318 (function and method decorator syntax)
break
answers.append(ans) • PEP 3129 (class decorator syntax)
return answers
• https://wiki.python.org/moin/PythonDecoratorLibrary
This is fine, as long as the body of the loop is fairly compact. Once it becomes more complicated, as • https://docs.python.org/dev/library/functools.html
often happens in real code, this becomes pretty unreadable. We could simplify this by using yield

7.2. Decorators 308 7.2. Decorators 309

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

exception, if thrown, is propagated. As with files, there’s often a natural operation to perform after the
• https://pypi.org/project/decorator object has been used and it is most convenient to have the support built in. With each release, Python
• Bruce Eckel provides support in more places:
– Decorators I: Introduction to Python Decorators • all file-like objects:
– Python Decorators II: Decorator Arguments – file ➥ automatically closed
– Python Decorators III: A Decorator-Based Build System – fileinput, tempfile
– bz2.BZ2File, gzip.GzipFile, tarfile.TarFile, zipfile.ZipFile
– ftplib, nntplib ➥ close connection
7.3 Context managers
• locks
A context manager is an object with __enter__ and __exit__ methods which can be used in the with
– multiprocessing.RLock ➥ lock and unlock
statement:
– multiprocessing.Semaphore
with manager as var:
do_something(var) – memoryview ➥ automatically release
• decimal.localcontext ➥ modify precision of computations temporarily
is in the simplest case equivalent to
• _winreg.PyHKEY ➥ open and close hive key
var = manager.__enter__()
• warnings.catch_warnings ➥ kill warnings temporarily
try:
do_something(var) • contextlib.closing ➥ the same as the example above, call close
finally:
• parallel programming
manager.__exit__()
– concurrent.futures.ThreadPoolExecutor ➥ invoke in parallel then kill thread pool
In other words, the context manager protocol defined in PEP 343 permits the extraction of the boring
– concurrent.futures.ProcessPoolExecutor ➥ invoke in parallel then kill process pool
part of a try..except..finally structure into a separate class leaving only the interesting do_something
block. – nogil ➥ solve the GIL problem temporarily (cython only :( )
1. The __enter__ method is called first. It can return a value which will be assigned to var. The
as-part is optional: if it isn’t present, the value returned by __enter__ is simply ignored. 7.3.1 Catching exceptions
2. The block of code underneath with is executed. Just like with try clauses, it can either execute When an exception is thrown in the with-block, it is passed as arguments to __exit__. Three arguments
successfully to the end, or it can break, continue or return, or it can throw an exception. Either are used, the same as returned by sys.exc_info(): type, value, traceback. When no exception is thrown,
way, after the block is finished, the __exit__ method is called. If an exception was thrown, None is used for all three arguments. The context manager can “swallow” the exception by returning a
the information about the exception is passed to __exit__, which is described below in the next true value from __exit__. Exceptions can be easily ignored, because if __exit__ doesn’t use return
subsection. In the normal case, exceptions can be ignored, just like in a finally clause, and will and just falls of the end, None is returned, a false value, and therefore the exception is rethrown after
be rethrown after __exit__ is finished. __exit__ is finished.
Let’s say we want to make sure that a file is closed immediately after we are done writing to it: The ability to catch exceptions opens interesting possibilities. A classic example comes from unit-tests
— we want to make sure that some code throws the right kind of exception:
>>> class closing(object):
... def __init__(self, obj): class assert_raises(object):
... self.obj = obj # based on pytest and unittest.TestCase
... def __enter__(self): def __init__(self, type):
... return self.obj self.type = type
... def __exit__(self, *args): def __enter__(self):
... self.obj.close() pass
>>> with closing(open('/tmp/file', 'w')) as f: def __exit__(self, type, value, traceback):
... f.write('the contents\n') if type is None:
raise AssertionError('exception expected')
Here we have made sure that the f.close() is called when the with block is exited. Since closing files is if issubclass(type, self.type):
such a common operation, the support for this is already present in the file class. It has an __exit__ return True # swallow the expected exception
method which calls close and can be used as a context manager itself: raise AssertionError('wrong exception type')

>>> with open('/tmp/file', 'a') as f: with assert_raises(KeyError):

... f.write('more contents\n') {}['foo']

The common use for try..finally is releasing resources. Various different cases are implemented
similarly: in the __enter__ phase the resource is acquired, in the __exit__ phase it is released, and the

7.3. Context managers 310 7.3. Context managers 311

 Scientific Python Lectures, Edition 2025.1rc0.dev0

7.3.2 Using generators to define context managers

When discussing generators, it was said that we prefer generators to iterators implemented as classes
because they are shorter, sweeter, and the state is stored as local, not instance, variables. On the other
hand, as described in Bidirectional communication, the flow of data between the generator and its caller
can be bidirectional. This includes exceptions, which can be thrown into the generator. We would like to

8
implement context managers as special generator functions. In fact, the generator protocol was designed
to support this use case.

@contextlib.contextmanager
def some_generator(<arguments>):
<setup>
try: CHAPTER
yield <value>
finally:
<cleanup>

The contextlib.contextmanager helper takes a generator and turns it into a context manager. The
generator has to obey some rules which are enforced by the wrapper function — most importantly it
must yield exactly once. The part before the yield is executed from __enter__, the block of code
protected by the context manager is executed when the generator is suspended in yield, and the rest
is executed in __exit__. If an exception is thrown, the interpreter hands it to the wrapper through
__exit__ arguments, and the wrapper function then throws it at the point of the yield statement. Advanced NumPy
Through the use of generators, the context manager is shorter and simpler.
Let’s rewrite the closing example as a generator:

@contextlib.contextmanager
def closing(obj):
try:
yield obj Author: Pauli Virtanen
finally:
obj.close() NumPy is at the base of Python’s scientific stack of tools. Its purpose to implement efficient operations
on many items in a block of memory. Understanding how it works in detail helps in making efficient use
Let’s rewrite the assert_raises example as a generator: of its flexibility, taking useful shortcuts.
This section covers:
@contextlib.contextmanager
def assert_raises(type): • Anatomy of NumPy arrays, and its consequences. Tips and tricks.
try:
• Universal functions: what, why, and what to do if you want a new one.
yield
except type: • Integration with other tools: NumPy offers several ways to wrap any data in an ndarray, without
return unnecessary copies.
except Exception as value:
• Recently added features, and what’s in them: PEP 3118 buffers, generalized ufuncs, . . .
raise AssertionError('wrong exception type')
else:
raise AssertionError('exception expected') Prerequisites

Here we use a decorator to turn generator functions into context managers! • NumPy
• Cython
• Pillow (Python imaging library, used in a couple of examples)

Chapter contents

• Life of ndarray
– It’s. . .

7.3. Context managers 312 313

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

– Block of memory
– Data types
– Indexing scheme: strides
– Findings in dissection
• Universal functions
– What they are?
– Exercise: building an ufunc from scratch
– Solution: building an ufunc from scratch typedef struct PyArrayObject {
PyObject_HEAD
– Generalized ufuncs
• Interoperability features /* Block of memory */
char *data;
– Sharing multidimensional, typed data
– The old buffer protocol /* Data type descriptor */
PyArray_Descr *descr;
– The old buffer protocol
– Array interface protocol /* Indexing scheme */
int nd;
• Array siblings: chararray, maskedarray
npy_intp *dimensions;
– chararray: vectorized string operations npy_intp *strides;
– masked_array missing data
/* Other stuff */
– recarray: purely convenience PyObject *base;
int flags;
• Summary
PyObject *weakreflist;
• Contributing to NumPy/SciPy } PyArrayObject;
– Why
– Reporting bugs 8.1.2 Block of memory
– Contributing to documentation >>> x = np.array([1, 2, 3], dtype=np.int32)
– Contributing features >>> x.data
<... at ...>
– How to help, in general >>> bytes(x.data)
b'\x01\x00\x00\x00\x02\x00\x00\x00\x03\x00\x00\x00'

Memory address of the data:

 Tip
>>> x.__array_interface__['data'][0]
In this section, NumPy will be imported as follows: 64803824
>>> import numpy as np
The whole __array_interface__:

>>> x.__array_interface__
8.1 Life of ndarray {'data': (..., False), 'strides': None, 'descr': [('', '<i4')], 'typestr': '<i4',
˓→'shape': (3,), 'version': 3}

8.1.1 It’s. . . Reminder: two ndarrays may share the same memory:
ndarray =
>>> x = np.array([1, 2, 3, 4])
block of memory + indexing scheme + data type descriptor >>> y = x[:-1]
• raw data >>> x[0] = 9
>>> y
• how to locate an element array([9, 2, 3])
• how to interpret an element

8.1. Life of ndarray 314 8.1. Life of ndarray 315

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Memory does not need to be owned by an ndarray:

chunk_id "RIFF"
>>> x = b'1234' chunk_size 4-byte unsigned little-endian integer
format "WAVE"
x is a string (in Python 3 a bytes), we can represent its data as an array of ints: fmt_id "fmt "
fmt_size 4-byte unsigned little-endian integer
>>> y = np.frombuffer(x, dtype=np.int8)
audio_fmt 2-byte unsigned little-endian integer
>>> y.data
num_channels 2-byte unsigned little-endian integer
<... at ...>
sample_rate 4-byte unsigned little-endian integer
>>> y.base is x
byte_rate 4-byte unsigned little-endian integer
True
block_align 2-byte unsigned little-endian integer
bits_per_sample 2-byte unsigned little-endian integer
>>> y.flags
data_id "data"
C_CONTIGUOUS : True
data_size 4-byte unsigned little-endian integer
F_CONTIGUOUS : True
OWNDATA : False
• 44-byte block of raw data (in the beginning of the file)
WRITEABLE : False
ALIGNED : True • . . . followed by data_size bytes of actual sound data.
WRITEBACKIFCOPY : False
The .wav file header as a NumPy structured data type:
The owndata and writeable flags indicate status of the memory block. >>> wav_header_dtype = np.dtype([
... ("chunk_id", (bytes, 4)), # flexible-sized scalar type, item size 4
... ("chunk_size", "<u4"), # little-endian unsigned 32-bit integer
ã See also
... ("format", "S4"), # 4-byte string
array interface ... ("fmt_id", "S4"),
... ("fmt_size", "<u4"),
... ("audio_fmt", "<u2"), #
8.1.3 Data types ... ("num_channels", "<u2"), # .. more of the same ...
... ("sample_rate", "<u4"), #
The descriptor ... ("byte_rate", "<u4"),
... ("block_align", "<u2"),
dtype describes a single item in the array:
... ("bits_per_sample", "<u2"),
... ("data_id", ("S1", (2, 2))), # sub-array, just for fun!
type scalar type of the data, one of: ... ("data_size", "u4"),
int8, int16, float64, et al. (fixed size) ... #
str, unicode, void (flexible size) ... # the sound data itself cannot be represented here:
itemsize size of the data block ... # it does not have a fixed size
byte- byte order: big-endian > / little-endian < / not applicable | ... ])
order
fields sub-dtypes, if it’s a structured data type
shape shape of the array, if it’s a sub-array ã See also

wavreader.py
>>> np.dtype(int).type
<class 'numpy.int64'>
>>> np.dtype(int).itemsize >>> wav_header_dtype['format']
8 dtype('S4')
>>> np.dtype(int).byteorder >>> wav_header_dtype.fields
'=' mappingproxy({'chunk_id': (dtype('S4'), 0), 'chunk_size': (dtype('uint32'), 4),
˓→'format': (dtype('S4'), 8), 'fmt_id': (dtype('S4'), 12), 'fmt_size': (dtype('uint32

˓→'), 16), 'audio_fmt': (dtype('uint16'), 20), 'num_channels': (dtype('uint16'), 22),

Example: reading .wav files ˓→'sample_rate': (dtype('uint32'), 24), 'byte_rate': (dtype('uint32'), 28), 'block_

˓→align': (dtype('uint16'), 32), 'bits_per_sample': (dtype('uint16'), 34), 'data_id':␣

The .wav file header:
˓→(dtype(('S1', (2, 2))), 36), 'data_size': (dtype('uint32'), 40)})

>>> wav_header_dtype.fields['format']
(dtype('S4'), 8)

• The first element is the sub-dtype in the structured data, corresponding to the name format

8.1. Life of ndarray 316 8.1. Life of ndarray 317

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• The second one is its offset (in bytes) from the beginning of the item Casting

• Casting in arithmetic, in nutshell:

Exercise
– only type (not value!) of operands matters
Mini-exercise, make a “sparse” dtype by using offsets, and only some of the fields:
– largest “safe” type able to represent both is picked
>>> wav_header_dtype = np.dtype(dict(
– scalars can “lose” to arrays in some situations
... names=['format', 'sample_rate', 'data_id'],
... offsets=[offset_1, offset_2, offset_3], # counted from start of structure in␣ • Casting in general copies data:
˓→bytes

... formats=list of dtypes for each of the fields, >>> x = np.array([1, 2, 3, 4], dtype=float)
... )) >>> x
array([1., 2., 3., 4.])
and use that to read the sample rate, and data_id (as sub-array). >>> y = x.astype(np.int8)
>>> y
array([1, 2, 3, 4], dtype=int8)
>>> f = open('data/test.wav', 'r')
>>> y + 1
>>> wav_header = np.fromfile(f, dtype=wav_header_dtype, count=1)
array([2, 3, 4, 5], dtype=int8)
>>> f.close()
>>> y + 256
>>> print(wav_header)
Traceback (most recent call last):
[ ('RIFF', 17402L, 'WAVE', 'fmt ', 16L, 1, 1, 16000L, 32000L, 2, 16, [['d', 'a'], ['t
File "<stdin>", line 1, in <module>
˓→', 'a']], 17366L)]
OverflowError: Python integer 256 out of bounds for int8
>>> wav_header['sample_rate']
>>> y + 256.0
array([16000], dtype=uint32)
array([257., 258., 259., 260.])
>>> y + np.array([256], dtype=np.int32)
Let’s try accessing the sub-array: array([257, 258, 259, 260], dtype=int32)
>>> wav_header['data_id']
array([[['d', 'a'], • Casting on setitem: dtype of the array is not changed on item assignment:
['t', 'a']]],
>>> y[:] = y + 1.5
dtype='|S1')
>>> y
>>> wav_header.shape
array([2, 3, 4, 5], dtype=int8)
(1,)
>>> wav_header['data_id'].shape
(1, 2, 2)
ò Note
When accessing sub-arrays, the dimensions get added to the end! Exact rules: see NumPy documentation

ò Note
Re-interpretation / viewing
There are existing modules such as wavfile, audiolab, etc. for loading sound data. . .
• Data block in memory (4 bytes)

Casting and re-interpretation/views 0x01 || 0x02 || 0x03 || 0x04

casting
– 4 of uint8, OR,
• on assignment
– 4 of int8, OR,
• on array construction
– 2 of int16, OR,
• on arithmetic
– 1 of int32, OR,
• etc.
– 1 of float32, OR,
• and manually: .astype(dtype)
– ...
data re-interpretation
How to switch from one to another?
• manually: .view(dtype)
1. Switch the dtype:

8.1. Life of ndarray 318 8.1. Life of ndarray 319

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

>>> x = np.array([1, 2, 3, 4], dtype=np.uint8) >>> y = ...

>>> x.dtype = "<i2"
>>> x >>> assert (y['r'] == 1).all()
array([ 513, 1027], dtype=int16) >>> assert (y['g'] == 2).all()
>>> 0x0201, 0x0403 >>> assert (y['b'] == 3).all()
(513, 1027) >>> assert (y['a'] == 4).all()

Solution
0x01 0x02 || 0x03 0x04
>>> y = x.view([('r', 'i1'),
... ('g', 'i1'),
... ('b', 'i1'),
ò Note
... ('a', 'i1')]
little-endian: least significant byte is on the left in memory ... )[:, :, 0]

2. Create a new view of type uint32, shorthand i4:

. Warning
>>> y = x.view("<i4")
>>> y Another two arrays, each occupying exactly 4 bytes of memory:
array([67305985], dtype=int32) >>> x = np.array([[1, 3], [2, 4]], dtype=np.uint8)
>>> 0x04030201 >>> x
67305985 array([[1, 3],
[2, 4]], dtype=uint8)
>>> y = x.transpose()
0x01 0x02 0x03 0x04 >>> y
array([[1, 2],
[3, 4]], dtype=uint8)
ò Note
We view the elements of x (1 byte each) as int16 (2 bytes each):
• .view() makes views, does not copy (or alter) the memory block >>> x.view(np.int16)
• only changes the dtype (and adjusts array shape): array([[ 769],
[1026]], dtype=int16)
>>> x[1] = 5
>>> y What is happening here? Take a look at the bytes stored in memory by x:
array([328193], dtype=int32) >>> x.tobytes()
>>> y.base is x b'\x01\x03\x02\x04'
True
The \x stands for heXadecimal, so what we are seeing is:
0x01 0x03 0x02 0x04
Mini-exercise: data re-interpretation
We ask NumPy to interpret these bytes as elements of dtype int16—each of which occupies two
bytes in memory. Therefore, 0x01 0x03 becomes the first uint16 and 0x02 0x04 the second.
ã See also
You may then expect to see 0x0103 (259, when converting from hexadecimal to decimal) as the first
view-colors.py result. But your computer likely stores most significant bytes first, and as such reads the number as
0x0301 or 769 (go on and type 0x0301 into your Python terminal to verify).
You have RGBA data in an array: We can do the same on a copy of y (why doesn’t it work on y directly?):

>>> x = np.zeros((10, 10, 4), dtype=np.int8) >>> y.copy().view(np.int16)

>>> x[:, :, 0] = 1 array([[ 513],
>>> x[:, :, 1] = 2 [1027]], dtype=int16)
>>> x[:, :, 2] = 3
>>> x[:, :, 3] = 4 Can you explain these numbers, 513 and 1027, as well as the output shape of the resulting array?

where the last three dimensions are the R, B, and G, and alpha channels.
How to make a (10, 10) structured array with field names ‘r’, ‘g’, ‘b’, ‘a’ without copying data?

8.1. Life of ndarray 320 8.1. Life of ndarray 321

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

8.1.4 Indexing scheme: strides – C: last dimensions vary fastest (= smaller strides)

Main point – F: first dimensions vary fastest

The question: shape = (𝑑1 , 𝑑2 , ..., 𝑑𝑛 )

strides = (𝑠1 , 𝑠2 , ..., 𝑠𝑛 )
>>> x = np.array([[1, 2, 3],
... [4, 5, 6], 𝑗 = 𝑑𝑗+1 𝑑𝑗+2 ...𝑑𝑛 × itemsize
𝑠𝐶
... [7, 8, 9]], dtype=np.int8) 𝑗 = 𝑑1 𝑑2 ...𝑑𝑗−1 × itemsize
𝑠𝐹
>>> x.tobytes('A')
b'\x01\x02\x03\x04\x05\x06\x07\x08\t'
ò Note
At which byte in ``x.data`` does the item ``x[1, 2]`` begin?
Now we can understand the behavior of .view():
The answer (in NumPy) >>> y = np.array([[1, 3], [2, 4]], dtype=np.uint8).transpose()
>>> x = y.copy()
• strides: the number of bytes to jump to find the next element
• 1 stride per dimension Transposition does not affect the memory layout of the data, only strides
>>> x.strides
>>> x.strides
(2, 1)
(3, 1)
>>> y.strides
>>> byte_offset = 3 * 1 + 1 * 2 # to find x[1, 2]
(1, 2)
>>> x.flat[byte_offset]
np.int8(6) >>> x.tobytes('A')
>>> x[1, 2] b'\x01\x02\x03\x04'
np.int8(6) >>> y.tobytes('A')
b'\x01\x03\x02\x04'
simple, flexible
• the results are different when interpreted as 2 of int16
C and Fortran order • .copy() creates new arrays in the C order (by default)

ò Note
ò Note
The Python built-in bytes returns bytes in C-order by default which can cause confusion when trying
to inspect memory layout. We use numpy.ndarray.tobytes() with order=A instead, which preserves In-place operations with views
the C or F ordering of the bytes in memory.
Prior to NumPy version 1.13, in-place operations with views could result in incorrect results for large
arrays. Since version 1.13, NumPy includes checks for memory overlap to guarantee that results are
>>> x = np.array([[1, 2, 3], consistent with the non in-place version (e.g. a = a + a.T produces the same result as a += a.T).
... [4, 5, 6]], dtype=np.int16, order='C') Note however that this may result in the data being copied (as if using a += a.T.copy()), ultimately
>>> x.strides resulting in more memory being used than might otherwise be expected for in-place operations!
(6, 2)
>>> x.tobytes('A')
b'\x01\x00\x02\x00\x03\x00\x04\x00\x05\x00\x06\x00' Slicing with integers

• Everything can be represented by changing only shape, strides, and possibly adjusting the data
• Need to jump 6 bytes to find the next row
pointer!
• Need to jump 2 bytes to find the next column
• Never makes copies of the data
>>> y = np.array(x, order='F')
>>> x = np.array([1, 2, 3, 4, 5, 6], dtype=np.int32)
>>> y.strides
>>> y = x[::-1]
(2, 4)
>>> y
>>> y.tobytes('A')
array([6, 5, 4, 3, 2, 1], dtype=int32)
b'\x01\x00\x04\x00\x02\x00\x05\x00\x03\x00\x06\x00'
>>> y.strides
(-4,)
• Need to jump 2 bytes to find the next row
• Need to jump 4 bytes to find the next column >>> y = x[2:]
>>> y.__array_interface__['data'][0] - x.__array_interface__['data'][0]
• Similarly to higher dimensions: (continues on next page)

8.1. Life of ndarray 322 8.1. Life of ndarray 323

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

ã See also
8
stride-fakedims.py
>>> x = np.zeros((10, 10, 10), dtype=float)
>>> x.strides
(800, 80, 8) Exercise
>>> x[::2,::3,::4].strides
array([1, 2, 3, 4], dtype=np.int8)
(1600, 240, 32)
-> array([[1, 2, 3, 4],
• Similarly, transposes never make copies (it just swaps strides):
[1, 2, 3, 4],
>>> x = np.zeros((10, 10, 10), dtype=float) [1, 2, 3, 4]], dtype=np.int8)
>>> x.strides
(800, 80, 8) using only as_strided.:
>>> x.T.strides
Hint: byte_offset = stride[0]*index[0] + stride[1]*index[1] + ...
(8, 80, 800)
Spoiler
But: not all reshaping operations can be represented by playing with strides:
Stride can also be 0 :
>>> a = np.arange(6, dtype=np.int8).reshape(3, 2)
>>> b = a.T >>> x = np.array([1, 2, 3, 4], dtype=np.int8)
>>> b.strides >>> y = as_strided(x, strides=(0, 1), shape=(3, 4))
(1, 2) >>> y
array([[1, 2, 3, 4],
So far, so good. However: [1, 2, 3, 4],
[1, 2, 3, 4]], dtype=int8)
>>> bytes(a.data) >>> y.base.base is x
b'\x00\x01\x02\x03\x04\x05' True
>>> b
array([[0, 2, 4],
[1, 3, 5]], dtype=int8) Broadcasting
>>> c = b.reshape(3*2)
• Doing something useful with it: outer product of [1, 2, 3, 4] and [5, 6, 7]
>>> c
array([0, 2, 4, 1, 3, 5], dtype=int8) >>> x = np.array([1, 2, 3, 4], dtype=np.int16)
>>> x2 = as_strided(x, strides=(0, 1*2), shape=(3, 4))
Here, there is no way to represent the array c given one stride and the block of memory for a. Therefore, >>> x2
the reshape operation needs to make a copy here. array([[1, 2, 3, 4],
[1, 2, 3, 4],
Example: fake dimensions with strides [1, 2, 3, 4]], dtype=int16)
Stride manipulation
>>> y = np.array([5, 6, 7], dtype=np.int16)
>>> from numpy.lib.stride_tricks import as_strided >>> y2 = as_strided(y, strides=(1*2, 0), shape=(3, 4))
>>> help(as_strided) >>> y2
Help on function as_strided in module numpy.lib.stride_tricks: array([[5, 5, 5, 5],
... [6, 6, 6, 6],
[7, 7, 7, 7]], dtype=int16)

. Warning >>> x2 * y2
array([[ 5, 10, 15, 20],
as_strided does not check that you stay inside the memory block bounds. . . [ 6, 12, 18, 24],
[ 7, 14, 21, 28]], dtype=int16)
>>> x = np.array([1, 2, 3, 4], dtype=np.int16)
>>> as_strided(x, strides=(2*2, ), shape=(2, )) . . . seems somehow familiar . . .
array([1, 3], dtype=int16)
>>> x[::2] >>> x = np.array([1, 2, 3, 4], dtype=np.int16)
array([1, 3], dtype=int16) >>> y = np.array([5, 6, 7], dtype=np.int16)
(continues on next page)

8.1. Life of ndarray 324 8.1. Life of ndarray 325

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

ã See also
>>> x[np.newaxis,:] * y[:,np.newaxis]
array([[ 5, 10, 15, 20], stride-diagonals.py
[ 6, 12, 18, 24],
[ 7, 14, 21, 28]], dtype=int16)
Challenge
• Internally, array broadcasting is indeed implemented using 0-strides. Compute the tensor trace:

More tricks: diagonals >>> x = np.arange(5*5*5*5).reshape(5, 5, 5, 5)

>>> s = 0
>>> for i in range(5):
ã See also ... for j in range(5):
... s += x[j, i, j, i]
stride-diagonals.py
by striding, and using sum() on the result.
Challenge
>>> y = as_strided(x, shape=(5, 5), strides=(TODO, TODO))
• Pick diagonal entries of the matrix: (assume C memory order): >>> s2 = ...
>>> assert s == s2
>>> x = np.array([[1, 2, 3],
... [4, 5, 6], Solution
... [7, 8, 9]], dtype=np.int32)
>>> y = as_strided(x, shape=(5, 5), strides=((5*5*5 + 5)*x.itemsize,
>>> x_diag = as_strided(x, shape=(3,), strides=(???,)) ... (5*5 + 1)*x.itemsize))
>>> s2 = y.sum()
• Pick the first super-diagonal entries [2, 6].
• And the sub-diagonals? CPU cache effects
(Hint to the last two: slicing first moves the point where striding Memory layout can affect performance:
starts from.)
In [1]: x = np.zeros((20000,))
Solution
Pick diagonals: In [2]: y = np.zeros((20000*67,))[::67]

>>> x_diag = as_strided(x, shape=(3, ), strides=((3+1)*x.itemsize, )) In [3]: x.shape, y.shape

>>> x_diag Out[3]: ((20000,), (20000,))
array([1, 5, 9], dtype=int32)
In [4]: %timeit x.sum()
Slice first, to adjust the data pointer: 5.27 us +- 7.15 ns per loop (mean +- std. dev. of 7 runs, 100,000 loops each)
>>> as_strided(x[0, 1:], shape=(2, ), strides=((3+1)*x.itemsize, )) In [5]: %timeit y.sum()
array([2, 6], dtype=int32) 17 us +- 24.6 ns per loop (mean +- std. dev. of 7 runs, 100,000 loops each)
>>> as_strided(x[1:, 0], shape=(2, ), strides=((3+1)*x.itemsize, )) In [6]: x.strides, y.strides
array([4, 8], dtype=int32) Out[6]: ((8,), (536,))

ò Note Smaller strides are faster?

Using np.diag
>>> y = np.diag(x, k=1)
>>> y
array([2, 6], dtype=int32)

However,
>>> y.flags.owndata
False
• CPU pulls data from main memory to its cache in blocks

8.1. Life of ndarray 326 8.1. Life of ndarray 327

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• If many array items consecutively operated on fit in a single block (small stride): (continued from previous page)

– ⇒ fewer transfers needed */

char *input_1 = (char*)args[0];
– ⇒ faster char *input_2 = (char*)args[1];
char *output = (char*)args[2];
int i;
ã See also

• numexpr is designed to mitigate cache effects when evaluating array expressions. for (i = 0; i < dimensions[0]; ++i) {
*output = elementwise_function(*input_1, *input_2);
• numba is a compiler for Python code, that is aware of numpy arrays. input_1 += steps[0];
input_2 += steps[1];
output += steps[2];
8.1.5 Findings in dissection }
}

2. The NumPy part, built by

char types[3]

types[0] = NPY_BYTE /* type of first input arg */

types[1] = NPY_BYTE /* type of second input arg */
types[2] = NPY_BYTE /* type of third input arg */

PyObject *python_ufunc = PyUFunc_FromFuncAndData(

ufunc_loop,
• memory block: may be shared, .base, .data
NULL,
• data type descriptor: structured data, sub-arrays, byte order, casting, viewing, .astype(), .view() types,
1, /* ntypes */
• strided indexing: strides, C/F-order, slicing w/ integers, as_strided, broadcasting, stride tricks,
2, /* num_inputs */
diag, CPU cache coherence
1, /* num_outputs */
identity_element,
8.2 Universal functions name,
docstring,
unused)
8.2.1 What they are?
• Ufunc performs and elementwise operation on all elements of an array. • A ufunc can also support multiple different input-output type combinations.
Examples:
Making it easier
np.add, np.subtract, scipy.special.*, ...
3. ufunc_loop is of very generic form, and NumPy provides pre-made ones
• Automatically support: broadcasting, casting, . . .
PyUfunc_f_f float elementwise_func(float input_1)
• The author of an ufunc only has to supply the elementwise operation, NumPy takes care of the
PyUfunc_ff_f float elementwise_func(float input_1, float input_2)
rest.
PyUfunc_d_d double elementwise_func(double input_1)
• The elementwise operation needs to be implemented in C (or, e.g., Cython) PyUfunc_dd_d double elementwise_func(double input_1, double input_2)
PyUfunc_D_D elementwise_func(npy_cdouble *input, npy_cdouble* output)
Parts of an Ufunc PyUfunc_DD_D elementwise_func(npy_cdouble *in1, npy_cdouble *in2, npy_cdouble*
out)
1. Provided by user
• Only elementwise_func needs to be supplied
void ufunc_loop(void **args, int *dimensions, int *steps, void *data)
{ • . . . except when your elementwise function is not in one of the above forms
/*
* int8 output = elementwise_function(int8 input_1, int8 input_2) 8.2.2 Exercise: building an ufunc from scratch
*
* This function must compute the ufunc for many values at once, The Mandelbrot fractal is defined by the iteration
* in the way shown below.
(continues on next page) 𝑧 ← 𝑧2 + 𝑐

8.2. Universal functions 328 8.2. Universal functions 329

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

where 𝑐 = 𝑥 + 𝑖𝑦 is a complex number. This iteration is repeated – if 𝑧 stays finite no matter how long (continued from previous page)
the iteration runs, 𝑐 belongs to the Mandelbrot set. #
• Make ufunc called mandel(z0, c) that computes:
cdef double complex z = z_in[0]
z = z0 cdef double complex c = c_in[0]
for k in range(iterations): cdef int k # the integer we use in the for loop
z = z*z + c
#
say, 100 iterations or until z.real**2 + z.imag**2 > 1000. Use it to determine which c are in # TODO: write the Mandelbrot iteration for one point here,
the Mandelbrot set. # as you would write it in Python.
#
• Our function is a simple one, so make use of the PyUFunc_* helpers.
# Say, use 100 as the maximum number of iterations, and 1000
• Write it in Cython # as the cutoff for z.real**2 + z.imag**2.
#

ã See also TODO: mandelbrot iteration should go here

mandel.pyx, mandelplot.py
# Return the answer for this point
z_out[0] = z
#
# Fix the parts marked by TODO
# # Boilerplate Cython definitions
#
# # Pulls definitions from the NumPy C headers.
# Compile this file by (Cython >= 0.12 required because of the complex vars) # -------------------------------------------
#
# cython mandel.pyx from numpy cimport import_array, import_ufunc
# python setup.py build_ext -i from numpy cimport (PyUFunc_FromFuncAndData,
# PyUFuncGenericFunction)
# and try it out with, in this directory, from numpy cimport NPY_CDOUBLE, NP_DOUBLE, NPY_LONG
#
# >>> import mandel # Import all pre-defined loop functions
# >>> mandel.mandel(0, 1 + 2j) # you won't need all of them - keep the relevant ones
#
# from numpy cimport (
PyUFunc_f_f_As_d_d,
# The elementwise function PyUFunc_d_d,
# ------------------------ PyUFunc_f_f,
PyUFunc_g_g,
cdef void mandel_single_point(double complex *z_in, PyUFunc_F_F_As_D_D,
double complex *c_in, PyUFunc_F_F,
double complex *z_out) nogil: PyUFunc_D_D,
# PyUFunc_G_G,
# The Mandelbrot iteration PyUFunc_ff_f_As_dd_d,
# PyUFunc_ff_f,
PyUFunc_dd_d,
# PyUFunc_gg_g,
# Some points of note: PyUFunc_FF_F_As_DD_D,
# PyUFunc_DD_D,
# - It's *NOT* allowed to call any Python functions here. PyUFunc_FF_F,
# PyUFunc_GG_G)
# The Ufunc loop runs with the Python Global Interpreter Lock released.
# Hence, the ``nogil``.
# # Required module initialization
# - And so all local variables must be declared with ``cdef`` # ------------------------------
#
# - Note also that this function receives *pointers* to the data import_array()
(continues on next page) (continues on next page)

8.2. Universal functions 330 8.2. Universal functions 331

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

import_ufunc() )

Reminder: some pre-made Ufunc loops:

# The actual ufunc declaration
# ----------------------------
PyUfunc_f_f float elementwise_func(float input_1)
PyUfunc_ff_f float elementwise_func(float input_1, float input_2)
cdef PyUFuncGenericFunction loop_func[1]
PyUfunc_d_d double elementwise_func(double input_1)
cdef char input_output_types[3]
PyUfunc_dd_d double elementwise_func(double input_1, double input_2)
cdef void *elementwise_funcs[1]
PyUfunc_D_D elementwise_func(complex_double *input, complex_double* output)
PyUfunc_DD_D elementwise_func(complex_double *in1, complex_double *in2,
#
complex_double* out)
# Reminder: some pre-made Ufunc loops:
#
# ================ ======================================================= Type codes:
# ``PyUfunc_f_f`` ``float elementwise_func(float input_1)``
# ``PyUfunc_ff_f`` ``float elementwise_func(float input_1, float input_2)`` NPY_BOOL, NPY_BYTE, NPY_UBYTE, NPY_SHORT, NPY_USHORT, NPY_INT, NPY_UINT,
# ``PyUfunc_d_d`` ``double elementwise_func(double input_1)`` NPY_LONG, NPY_ULONG, NPY_LONGLONG, NPY_ULONGLONG, NPY_FLOAT, NPY_DOUBLE,
# ``PyUfunc_dd_d`` ``double elementwise_func(double input_1, double input_2)`` NPY_LONGDOUBLE, NPY_CFLOAT, NPY_CDOUBLE, NPY_CLONGDOUBLE, NPY_DATETIME,
# ``PyUfunc_D_D`` ``elementwise_func(complex_double *input, complex_double* complex_ NPY_TIMEDELTA, NPY_OBJECT, NPY_STRING, NPY_UNICODE, NPY_VOID
˓→double)``

# ``PyUfunc_DD_D`` ``elementwise_func(complex_double *in1, complex_double *in2,␣

8.2.3 Solution: building an ufunc from scratch
˓→complex_double* out)``

# ================ ======================================================= # The elementwise function

# # ------------------------
# The full list is above.
# cdef void mandel_single_point(double complex *z_in,
# double complex *c_in,
# Type codes: double complex *z_out) noexcept nogil:
# #
# NPY_BOOL, NPY_BYTE, NPY_UBYTE, NPY_SHORT, NPY_USHORT, NPY_INT, NPY_UINT, # The Mandelbrot iteration
# NPY_LONG, NPY_ULONG, NPY_LONGLONG, NPY_ULONGLONG, NPY_FLOAT, NPY_DOUBLE, #
# NPY_LONGDOUBLE, NPY_CFLOAT, NPY_CDOUBLE, NPY_CLONGDOUBLE, NPY_DATETIME,
# NPY_TIMEDELTA, NPY_OBJECT, NPY_STRING, NPY_UNICODE, NPY_VOID #
# # Some points of note:
#
loop_func[0] = ... TODO: suitable PyUFunc_* ... # - It's *NOT* allowed to call any Python functions here.
input_output_types[0] = ... TODO ... #
... TODO: fill in rest of input_output_types ... # The Ufunc loop runs with the Python Global Interpreter Lock released.
# Hence, the ``nogil``.
# This thing is passed as the ``data`` parameter for the generic #
# PyUFunc_* loop, to let it know which function it should call. # - And so all local variables must be declared with ``cdef``
elementwise_funcs[0] = <void*>mandel_single_point #
# - Note also that this function receives *pointers* to the data;
# Construct the ufunc: # the "traditional" solution to passing complex variables around
#
mandel = PyUFunc_FromFuncAndData(
loop_func, cdef double complex z = z_in[0]
elementwise_funcs, cdef double complex c = c_in[0]
input_output_types, cdef int k # the integer we use in the for loop
1, # number of supported input types
TODO, # number of input args # Straightforward iteration
TODO, # number of output args
0, # `identity` element, never mind this for k in range(100):
"mandel", # function name z = z*z + c
"mandel(z, c) -> computes z*z + c", # docstring if z.real**2 + z.imag**2 > 1000:
0 # unused break
(continues on next page) (continues on next page)

8.2. Universal functions 332 8.2. Universal functions 333

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

# Return the answer for this point """

z_out[0] = z
import numpy as np
import mandel
# Boilerplate Cython definitions
# x = np.linspace(-1.7, 0.6, 1000)
# Pulls definitions from the NumPy C headers. y = np.linspace(-1.4, 1.4, 1000)
# ------------------------------------------- c = x[None, :] + 1j * y[:, None]
z = mandel.mandel(c, c)
from numpy cimport import_array, import_ufunc
from numpy cimport (PyUFunc_FromFuncAndData, import matplotlib.pyplot as plt
PyUFuncGenericFunction)
from numpy cimport NPY_CDOUBLE plt.imshow(abs(z) ** 2 < 1000, extent=[-1.7, 0.6, -1.4, 1.4])
from numpy cimport PyUFunc_DD_D plt.gray()
plt.show()
# Required module initialization
# ------------------------------

import_array()
import_ufunc()

# The actual ufunc declaration

# ----------------------------

cdef PyUFuncGenericFunction loop_func[1]

cdef char input_output_types[3]
cdef void *elementwise_funcs[1]

loop_func[0] = PyUFunc_DD_D

input_output_types[0] = NPY_CDOUBLE
input_output_types[1] = NPY_CDOUBLE ò Note
input_output_types[2] = NPY_CDOUBLE
Most of the boilerplate could be automated by these Cython modules:
elementwise_funcs[0] = <void*>mandel_single_point https://github.com/cython/cython/wiki/MarkLodato-CreatingUfuncs

mandel = PyUFunc_FromFuncAndData(
loop_func, Several accepted input types
elementwise_funcs,
input_output_types, E.g. supporting both single- and double-precision versions
1, # number of supported input types cdef void mandel_single_point(double complex *z_in,
2, # number of input args double complex *c_in,
1, # number of output args double complex *z_out) nogil:
0, # `identity` element, never mind this ...
"mandel", # function name
"mandel(z, c) -> computes iterated z*z + c", # docstring cdef void mandel_single_point_singleprec(float complex *z_in,
0 # unused float complex *c_in,
) float complex *z_out) nogil:
...
"""
Plot Mandelbrot cdef PyUFuncGenericFunction loop_funcs[2]
================ cdef char input_output_types[3*2]
cdef void *elementwise_funcs[1*2]
Plot the Mandelbrot ensemble.
(continues on next page) (continues on next page)

8.2. Universal functions 334 8.2. Universal functions 335

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) Status in NumPy

loop_funcs[0] = PyUFunc_DD_D • g-ufuncs are in NumPy already . . .
input_output_types[0] = NPY_CDOUBLE
input_output_types[1] = NPY_CDOUBLE • new ones can be created with PyUFunc_FromFuncAndDataAndSignature
input_output_types[2] = NPY_CDOUBLE • most linear-algebra functions are implemented as g-ufuncs to enable working with stacked arrays:
elementwise_funcs[0] = <void*>mandel_single_point
>>> import numpy as np
loop_funcs[1] = PyUFunc_FF_F >>> rng = np.random.default_rng(27446968)
input_output_types[3] = NPY_CFLOAT >>> np.linalg.det(rng.random((3, 5, 5)))
input_output_types[4] = NPY_CFLOAT array([ 0.01829761, -0.0077266 , -0.05336566])
input_output_types[5] = NPY_CFLOAT >>> np.linalg._umath_linalg.det.signature
elementwise_funcs[1] = <void*>mandel_single_point_singleprec '(m,m)->()'

mandel = PyUFunc_FromFuncAndData( • matrix multiplication this way could be useful for operating on many small matrices at once
loop_func,
elementwise_funcs, • Also see tensordot and einsum
input_output_types,
2, # number of supported input types <---------------- Generalized ufunc loop
2, # number of input args Matrix multiplication (m,n),(n,p) -> (m,p)
1, # number of output args
0, # `identity` element, never mind this void gufunc_loop(void **args, int *dimensions, int *steps, void *data)
"mandel", # function name {
"mandel(z, c) -> computes iterated z*z + c", # docstring char *input_1 = (char*)args[0]; /* these are as previously */
0 # unused char *input_2 = (char*)args[1];
) char *output = (char*)args[2];

int input_1_stride_m = steps[3]; /* strides for the core dimensions */

8.2.4 Generalized ufuncs int input_1_stride_n = steps[4]; /* are added after the non-core */
ufunc int input_2_strides_n = steps[5]; /* steps */
int input_2_strides_p = steps[6];
output = elementwise_function(input) int output_strides_n = steps[7];
Both output and input can be a single array element only. int output_strides_p = steps[8];

generalized ufunc int m = dimension[1]; /* core dimensions are added after */

output and input can be arrays with a fixed number of dimensions int n = dimension[2]; /* the main dimension; order as in */
int p = dimension[3]; /* signature */
For example, matrix trace (sum of diag elements):
int i;
input shape = (n, n)
output shape = () i.e. scalar for (i = 0; i < dimensions[0]; ++i) {
matmul_for_strided_matrices(input_1, input_2, output,
(n, n) -> () strides for each array...);

Matrix product: input_1 += steps[0];

input_1 shape = (m, n) input_2 += steps[1];
input_2 shape = (n, p) output += steps[2];
output shape = (m, p) }
}
(m, n), (n, p) -> (m, p)

• This is called the “signature” of the generalized ufunc 8.3 Interoperability features
• The dimensions on which the g-ufunc acts, are “core dimensions”
8.3.1 Sharing multidimensional, typed data
Suppose you
1. Write a library than handles (multidimensional) binary data,
2. Want to make it easy to manipulate the data with NumPy, or whatever other library,

8.2. Universal functions 336 8.3. Interoperability features 337

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

3. . . . but would not like to have NumPy as a dependency. (continued from previous page)

Currently, 3 solutions: img = Image.frombuffer("RGBA", (200, 200), data)

img.save("test.png")
1. the “old” buffer interface
2. the array interface # Modify the original data, and save again.

3. the “new” buffer interface (PEP 3118) x[:, :, 1] = 255

img.save("test2.png")
8.3.2 The old buffer protocol
• Only 1-D buffers
• No data type information
• C-level interface; PyBufferProcs tp_as_buffer in the type object
• But it’s integrated into Python (e.g. strings support it)
Mini-exercise using Pillow (Python Imaging Library):

ã See also

pilbuffer.py

>>> from PIL import Image

>>> data = np.zeros((200, 200, 4), dtype=np.uint8)
>>> data[:, :] = [255, 0, 0, 255] # Red
>>> # In PIL, RGBA images consist of 32-bit integers whose bytes are [RR,GG,BB,AA]
>>> data = data.view(np.int32).squeeze()
>>> img = Image.frombuffer("RGBA", (200, 200), data, "raw", "RGBA", 0, 1)
>>> img.save('test.png')

Q:
Check what happens if data is now modified, and img saved again.

8.3.3 The old buffer protocol

""" 8.3.4 Array interface protocol
From buffer • Multidimensional buffers
============
• Data type information present
Show how to exchange data between numpy and a library that only knows • NumPy-specific approach; slowly deprecated (but not going away)
the buffer interface.
""" • Not integrated in Python otherwise

import numpy as np ã See also

from PIL import Image
Documentation: https://numpy.org/doc/stable/reference/arrays.interface.html
# Let's make a sample image, RGBA format

x = np.zeros((200, 200, 4), dtype=np.uint8) >>> x = np.array([[1, 2], [3, 4]])

>>> x.__array_interface__
x[:, :, 0] = 255 # red {'data': (171694552, False), # memory address of data, is readonly?
x[:, :, 3] = 255 # opaque 'descr': [('', '<i4')], # data type descriptor
'typestr': '<i4', # same, in another form
data = x.view(np.int32) # Check that you understand why this is OK! 'strides': None, # strides; or None if in C-order
'shape': (2, 2),
(continues on next page) (continues on next page)

8.3. Interoperability features 338 8.3. Interoperability features 339

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

. Warning
'version': 3,
} Not all NumPy functions respect masks, for instance np.dot, so check the return types.

::
The masked_array returns a view to the original array:
>>> from PIL import Image
>>> img = Image.open('data/test.png') >>> mx[1] = 9
>>> img.__array_interface__ >>> x
{'version': 3, array([ 1, 9, 3, -99, 5])
'data': ...,
'shape': (200, 200, 4),
The mask
'typestr': '|u1'}
>>> x = np.asarray(img) You can modify the mask by assigning:
>>> x.shape
(200, 200, 4) >>> mx[1] = np.ma.masked
>>> mx
masked_array(data=[1, --, 3, --, 5],
ò Note mask=[False, True, False, True, False],
fill_value=999999)
A more C-friendly variant of the array interface is also defined.
The mask is cleared on assignment:

>>> mx[1] = 9
8.4 Array siblings: chararray, maskedarray >>> mx
masked_array(data=[1, 9, 3, --, 5],
8.4.1 chararray: vectorized string operations mask=[False, False, False, True, False],
fill_value=999999)
>>> x = np.char.asarray(['a', ' bbb', ' ccc'])
>>> x The mask is also available directly:
chararray(['a', ' bbb', ' ccc'], dtype='<U5')
>>> x.upper() >>> mx.mask
chararray(['A', ' BBB', ' CCC'], dtype='<U5') array([False, False, False, True, False])

The masked entries can be filled with a given value to get an usual array back:
8.4.2 masked_array missing data
>>> x2 = mx.filled(-1)
Masked arrays are arrays that may have missing or invalid entries.
>>> x2
For example, suppose we have an array where the fourth entry is invalid: array([ 1, 9, 3, -1, 5])

>>> x = np.array([1, 2, 3, -99, 5]) The mask can also be cleared:

One way to describe this is to create a masked array: >>> mx.mask = np.ma.nomask
>>> mx
>>> mx = np.ma.masked_array(x, mask=[0, 0, 0, 1, 0]) masked_array(data=[1, 9, 3, -99, 5],
>>> mx mask=[False, False, False, False, False],
masked_array(data=[1, 2, 3, --, 5], fill_value=999999)
mask=[False, False, False, True, False],
fill_value=999999)
Domain-aware functions
Masked mean ignores masked data:
The masked array package also contains domain-aware functions:
>>> mx.mean()
>>> np.ma.log(np.array([1, 2, -1, -2, 3, -5]))
np.float64(2.75)
masked_array(data=[0.0, 0.693147180559..., --, --, 1.098612288668..., --],
>>> np.mean(mx)
mask=[False, False, True, True, False, True],
np.float64(2.75)
fill_value=1e+20)

8.4. Array siblings: chararray, maskedarray 340 8.4. Array siblings: chararray, maskedarray 341
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

ò Note
array([b'a', b'b'], dtype='|S1')
Streamlined and more seamless support for dealing with missing data in arrays is making its way >>> arr2.y
into NumPy 1.7. Stay tuned! array([1, 2])

Example: Masked statistics 8.5 Summary

Canadian rangers were distracted when counting hares and lynxes in 1903-1910 and 1917-1918, and • Anatomy of the ndarray: data, dtype, strides.
got the numbers are wrong. (Carrot farmers stayed alert, though.) Compute the mean populations
• Universal functions: elementwise operations, how to make new ones
over time, ignoring the invalid numbers.
• Ndarray subclasses
>>> data = np.loadtxt('data/populations.txt')
>>> populations = np.ma.masked_array(data[:,1:]) • Various buffer interfaces for integration with other tools
>>> year = data[:, 0]
• Recent additions: PEP 3118, generalized ufuncs
>>> bad_years = (((year >= 1903) & (year <= 1910))
... | ((year >= 1917) & (year <= 1918))) 8.6 Contributing to NumPy/SciPy
>>> # '&' means 'and' and '|' means 'or'
>>> populations[bad_years, 0] = np.ma.masked Get this tutorial: https://www.euroscipy.org/talk/882
>>> populations[bad_years, 1] = np.ma.masked

>>> populations.mean(axis=0)
8.6.1 Why
masked_array(data=[40472.72727272727, 18627.272727272728, 42400.0], • “There’s a bug?”
mask=[False, False, False],
fill_value=1e+20) • “I don’t understand what this is supposed to do?”
• “I have this fancy code. Would you like to have it?”
>>> populations.std(axis=0)
masked_array(data=[21087.656489006717, 15625.799814240254, 3322.5062255844787], • “I’d like to help! What can I do?”
mask=[False, False, False],
fill_value=1e+20) 8.6.2 Reporting bugs
Note that Matplotlib knows about masked arrays: • Bug tracker (prefer this)

>>> plt.plot(year, populations, 'o-') – https://github.com/numpy/numpy/issues

[<matplotlib.lines.Line2D object at ...>, ...] – https://github.com/scipy/scipy/issues
– Click the “Sign up” link to get an account
• Mailing lists (https://numpy.org/community/)
– If you’re unsure
– No replies in a week or so? Just file a bug ticket.

Good bug report

Title: numpy.random.permutations fails for non-integer arguments

I'm trying to generate random permutations, using numpy.random.permutations

When calling numpy.random.permutation with non-integer arguments

it fails with a cryptic error message::

>>> rng.permutation(12)
8.4.3 recarray: purely convenience array([ 2, 6, 4, 1, 8, 11, 10, 5, 9, 3, 7, 0])
>>> rng.permutation(12.) #doctest: +SKIP
>>> arr = np.array([('a', 1), ('b', 2)], dtype=[('x', 'S1'), ('y', int)]) Traceback (most recent call last):
>>> arr2 = arr.view(np.recarray) File "<stdin>", line 1, in <module>
>>> arr2.x File "_generator.pyx", line 4844, in numpy.random._generator.Generator.permutation
(continues on next page) (continues on next page)

8.4. Array siblings: chararray, maskedarray 342 8.5. Summary 343

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

numpy.exceptions.AxisError: axis 0 is out of bounds for array of dimension 0 Cheers,
N. N.
This also happens with long arguments, and so
np.random.permutation(X.shape[0]) where X is an array fails on 64 • Check the style guide:
bit windows (where shape is a tuple of longs).
– https://numpy.org/doc/stable/
It would be great if it could cast to integer or at least raise a – Don’t be intimidated; to fix a small thing, just fix it
proper error for non-integer types.
• Edit
I'm using NumPy 1.4.1, built from the official tarball, on Windows 2. Edit sources and send patches (as for bugs)
64 with Visual studio 2008, on Python.org 64-bit Python.
3. Complain on the mailing list
0. What are you trying to do?
8.6.4 Contributing features
1. Small code snippet reproducing the bug (if possible)
The contribution of features is documented on https://numpy.org/doc/stable/dev/
• What actually happens
• What you’d expect 8.6.5 How to help, in general
2. Platform (Windows / Linux / OSX, 32/64 bits, x86/PPC, . . . ) • Bug fixes always welcome!
3. Version of NumPy/SciPy – What irks you most
>>> print(np.__version__) – Browse the tracker
2...
• Documentation work
Check that the following is what you expect – API docs: improvements to docstrings

>>> print(np.__file__) ∗ Know some SciPy module well?

/... – User guide

In case you have old/broken NumPy installations lying around. ∗ https://numpy.org/doc/stable/user/

If unsure, try to remove existing NumPy installations, and reinstall. . . • Ask on communication channels:
– numpy-discussion list
8.6.3 Contributing to documentation – scipy-dev list
1. Documentation editor
• https://numpy.org/doc/stable/
• Registration
– Register an account
– Subscribe to scipy-dev mailing list (subscribers-only)
– Problem with mailing lists: you get mail
∗ But: you can turn mail delivery off
∗ “change your subscription options”, at the bottom of
https://mail.python.org/mailman3/lists/scipy-dev.python.org/
– Send a mail @ scipy-dev mailing list; ask for activation:

To: scipy-dev@scipy.org

Hi,

I'd like to edit NumPy/SciPy docstrings. My account is XXXXX

(continues on next page)

8.6. Contributing to NumPy/SciPy 344 8.6. Contributing to NumPy/SciPy 345

 Scientific Python Lectures, Edition 2025.1rc0.dev0

– Debugger commands and interaction

• Debugging segmentation faults using gdb

9
9.1 Avoiding bugs
9.1.1 Coding best practices to avoid getting in trouble

CHAPTER Brian Kernighan

“Everyone knows that debugging is twice as hard as writing a program in the first place. So if you’re
as clever as you can be when you write it, how will you ever debug it?”

• We all write buggy code. Accept it. Deal with it.

• Write your code with testing and debugging in mind.
• Keep It Simple, Stupid (KISS).

Debugging code – What is the simplest thing that could possibly work?
• Don’t Repeat Yourself (DRY).
– Every piece of knowledge must have a single, unambiguous, authoritative representation within
a system.
– Constants, algorithms, etc. . .
• Try to limit interdependencies of your code. (Loose Coupling)
Author: Gaël Varoquaux • Give your variables, functions and modules meaningful names (not mathematics names)
This section explores tools to understand better your code base: debugging, to find and fix bugs.
It is not specific to the scientific Python community, but the strategies that we will employ are tailored 9.1.2 pyflakes: fast static analysis
to its needs. They are several static analysis tools in Python; to name a few:
• pylint
Prerequisites
• pychecker
• NumPy • pyflakes
• IPython • flake8
• nosetests Here we focus on pyflakes, which is the simplest tool.
• pyflakes • Fast, simple
• gdb for the C-debugging part. • Detects syntax errors, missing imports, typos on names.
Another good recommendation is the flake8 tool which is a combination of pyflakes and pep8. Thus, in
addition to the types of errors that pyflakes catches, flake8 detects violations of the recommendation in
Chapter contents
PEP8 style guide.
• Avoiding bugs Integrating pyflakes (or flake8) in your editor or IDE is highly recommended, it does yield productivity
gains.
– Coding best practices to avoid getting in trouble
– pyflakes: fast static analysis Running pyflakes on the current edited file
• Debugging workflow You can bind a key to run pyflakes in the current buffer.
• Using the Python debugger • In kate Menu: ‘settings -> configure kate
– Invoking the debugger – In plugins enable ‘external tools’

346 9.1. Avoiding bugs 347

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

– In external Tools’, add pyflakes:

kdialog --title "pyflakes %f ilename" --msgbox "$(pyflakes %f ilename)"

• In TextMate
Menu: TextMate -> Preferences -> Advanced -> Shell variables, add a shell variable:

TM_PYCHECKER = /Library/Frameworks/Python.framework/Versions/Current/bin/pyflakes

Then Ctrl-Shift-V is binded to a pyflakes report

• In vim In your .vimrc (binds F5 to pyflakes):

autocmd FileType python let &mp = 'echo "*** running % ***" ; pyflakes %' • In emacs
autocmd FileType tex,mp,rst,python imap <Esc>[15~ <C-O>:make!^M Use the flymake mode with pyflakes, documented on https://www.emacswiki.org/emacs/FlyMake
autocmd FileType tex,mp,rst,python map <Esc>[15~ :make!^M and included in Emacs 26 and more recent. To activate it, use M-x (meta-key then x) and enter
autocmd FileType tex,mp,rst,python set autowrite flymake-mode at the prompt. To enable it automatically when opening a Python file, add the
following line to your .emacs file:
• In emacs In your .emacs (binds F5 to pyflakes):
(add-hook 'python-mode-hook '(lambda () (flymake-mode)))
(defun pyflakes-thisfile () (interactive)
(compile (format "pyflakes %s " (buffer-file-name)))
)
9.2 Debugging workflow
(define-minor-mode pyflakes-mode
If you do have a non trivial bug, this is when debugging strategies kick in. There is no silver bullet. Yet,
"Toggle pyflakes mode.
strategies help:
With no argument, this command toggles the mode.
Non-null prefix argument turns on the mode. For debugging a given problem, the favorable situation is when the problem is
Null prefix argument turns off the mode." isolated in a small number of lines of code, outside framework or application
;; The initial value. code, with short modify-run-fail cycles
nil
1. Make it fail reliably. Find a test case that makes the code fail every time.
;; The indicator for the mode line.
" Pyflakes" 2. Divide and Conquer. Once you have a failing test case, isolate the failing code.
;; The minor mode bindings.
• Which module.
'( ([f5] . pyflakes-thisfile) )
) • Which function.
• Which line of code.
(add-hook 'python-mode-hook (lambda () (pyflakes-mode t)))
=> isolate a small reproducible failure: a test case

A type-as-go spell-checker like integration 3. Change one thing at a time and re-run the failing test case.

• In vim 4. Use the debugger to understand what is going wrong.

– Use the pyflakes.vim plugin: 5. Take notes and be patient. It may take a while.

1. download the zip file from https://www.vim.org/scripts/script.php?script_id=2441

ò Note
2. extract the files in ~/.vim/ftplugin/python
3. make sure your vimrc has filetype plugin indent on Once you have gone through this process: isolated a tight piece of code reproducing the bug and fix
the bug using this piece of code, add the corresponding code to your test suite.

9.3 Using the Python debugger

The python debugger, pdb: https://docs.python.org/3/library/pdb.html, allows you to inspect your code
interactively.
– Alternatively: use the syntastic plugin. This can be configured to use flake8 too and also
handles on-the-fly checking for many other languages. Specifically it allows you to:
• View the source code.

9.1. Avoiding bugs 348 9.2. Debugging workflow 349

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• Walk up and down the call stack. (continued from previous page)

• Inspect values of variables. 5 lst = list("foobar")

----> 6 print(lst[len(lst)])
• Modify values of variables. 7
• Set breakpoints. 8
9 if __name__ == "__main__":
10 index_error()
print
ipdb> len(lst)
Yes, print statements do work as a debugging tool. However to inspect runtime, it is often more 6
efficient to use the debugger. ipdb> print(lst[len(lst) - 1])
r
ipdb> quit
9.3.1 Invoking the debugger
Ways to launch the debugger:
Post-mortem debugging without IPython
1. Postmortem, launch debugger after module errors.
2. Launch the module with the debugger. In some situations you cannot use IPython, for instance to debug a script that wants to be called
from the command line. In this case, you can call the script with python -m pdb script.py:
3. Call the debugger inside the module
$ python -m pdb index_error.py
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/index_error.py(1)
Postmortem
˓→<module>()

Situation: You’re working in IPython and you get a traceback. -> """Small snippet to raise an IndexError."""
(Pdb) continue
Here we debug the file index_error.py. When running it, an IndexError is raised. Type %debug and
Traceback (most recent call last):
drop into the debugger.
File "/usr/lib64/python3.11/pdb.py", line 1793, in main
In [1]: %run index_error.py pdb._run(target)
--------------------------------------------------------------------------- File "/usr/lib64/python3.11/pdb.py", line 1659, in _run
IndexError Traceback (most recent call last) self.run(target.code)
File ~/src/scientific-python-lectures/advanced/debugging/index_error.py:10 File "/usr/lib64/python3.11/bdb.py", line 600, in run
6 print(lst[len(lst)]) exec(cmd, globals, locals)
9 if __name__ == "__main__": File "<string>", line 1, in <module>
---> 10 index_error() File "/home/jarrod/src/scientific-python-lectures/advanced/debugging/index_error.
˓→py", line 10, in <module>

File ~/src/scientific-python-lectures/advanced/debugging/index_error.py:6, in index_ index_error()

˓→error()
File "/home/jarrod/src/scientific-python-lectures/advanced/debugging/index_error.
˓→py", line 6, in index_error
4 def index_error():
5 lst = list("foobar") print(lst[len(lst)])
----> 6 print(lst[len(lst)]) ~~~^^^^^^^^^^
IndexError: list index out of range
IndexError: list index out of range Uncaught exception. Entering post mortem debugging
Running 'cont' or 'step' will restart the program
In [2]: %debug > /home/jarrod/src/scientific-python-lectures/advanced/debugging/index_error.
˓→py(6)index_error()
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/index_error.
˓→py(6)index_error()
-> print(lst[len(lst)])
4 def index_error(): (Pdb)
5 lst = list("foobar")
----> 6 print(lst[len(lst)])
7 Step-by-step execution
8
Situation: You believe a bug exists in a module but are not sure where.
ipdb> list For instance we are trying to debug wiener_filtering.py. Indeed the code runs, but the filtering does
1 """Small snippet to raise an IndexError.""" not work well.
2
3 • Run the script in IPython with the debugger using %run -d wiener_filtering.py :
4 def index_error():
(continues on next page)

9.3. Using the Python debugger 350 9.3. Using the Python debugger 351
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

In [1]: %run -d wiener_filtering.py
*** Blank or comment 32 for i in range(3):
*** Blank or comment 33 res = noisy_img - denoised_img
*** Blank or comment
NOTE: Enter 'c' at the ipdb> prompt to continue execution. • Step a few lines and explore the local variables:
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/wiener_
ipdb> n
˓→filtering.py(1)<module>()
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/wiener_
----> 1 """Wiener filtering a noisy raccoon face: this module is buggy"""
˓→filtering.py(32)iterated_wiener()
2
30 denoised_img = local_mean(noisy_img, size=size)
3 import numpy as np
31 l_var = local_var(noisy_img, size=size)
4 import scipy as sp
---> 32 for i in range(3):
5 import matplotlib.pyplot as plt
33 res = noisy_img - denoised_img
34 noise = (res**2).sum() / res.size
• Set a break point at line 29 using b 29:

ipdb> n ipdb> print(l_var)

> /home/jarrod/src/scientific-python-lectures/advanced/debugging/wiener_ [[2571 2782 3474 ... 3008 2922 3141]
˓→filtering.py(3)<module>()
[2105 708 475 ... 469 354 2884]
1 """Wiener filtering a noisy raccoon face: this module is buggy""" [1697 420 645 ... 273 236 2517]
2 ...
----> 3 import numpy as np [2437 345 432 ... 413 387 4188]
4 import scipy as sp [2598 179 247 ... 367 441 3909]
5 import matplotlib.pyplot as plt [2808 2525 3117 ... 4413 4454 4385]]
ipdb> print(l_var.min())
ipdb> b 29 0
Breakpoint 1 at /home/jarrod/src/scientific-python-lectures/advanced/debugging/
˓→wiener_filtering.py:29
Oh dear, nothing but integers, and 0 variation. Here is our bug, we are doing integer arithmetic.

• Continue execution to next breakpoint with c(ont(inue)): Raising exception on numerical errors
ipdb> c When we run the wiener_filtering.py file, the following warnings are raised:
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/wiener_
˓→filtering.py(29)iterated_wiener()
In [2]: %run wiener_filtering.py
27 Do not use this: this is crappy code to demo bugs! /home/jarrod/src/scientific-python-lectures/advanced/debugging/wiener_filtering.
˓→py:35: RuntimeWarning: divide by zero encountered in divide
28 """
1--> 29 noisy_img = noisy_img noise_level = 1 - noise / l_var
30 denoised_img = local_mean(noisy_img, size=size)
We can turn these warnings in exception, which enables us to do post-mortem debugging on them,
31 l_var = local_var(noisy_img, size=size)
and find our problem more quickly:
• Step into code with n(ext) and s(tep): next jumps to the next statement in the current execution In [3]: np.seterr(all='raise')
context, while step will go across execution contexts, i.e. enable exploring inside function calls: Out[3]: {'divide': 'warn', 'over': 'warn', 'under': 'ignore', 'invalid': 'warn'}

ipdb> s In [4]: %run wiener_filtering.py

> /home/jarrod/src/scientific-python-lectures/advanced/debugging/wiener_ ---------------------------------------------------------------------------
˓→filtering.py(30)iterated_wiener() FloatingPointError Traceback (most recent call last)
28 """ File ~/src/scientific-python-lectures/advanced/debugging/wiener_filtering.py:52
1 29 noisy_img = noisy_img 49 plt.matshow(face[cut], cmap=plt.cm.gray)
---> 30 denoised_img = local_mean(noisy_img, size=size) 50 plt.matshow(noisy_face[cut], cmap=plt.cm.gray)
31 l_var = local_var(noisy_img, size=size) ---> 52 denoised_face = iterated_wiener(noisy_face)
32 for i in range(3): 53 plt.matshow(denoised_face[cut], cmap=plt.cm.gray)
55 plt.show()
ipdb> n
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/wiener_ File ~/src/scientific-python-lectures/advanced/debugging/wiener_filtering.py:35,␣
˓→filtering.py(31)iterated_wiener() ˓→in iterated_wiener(noisy_img, size)

1 29 noisy_img = noisy_img 33 res = noisy_img - denoised_img

30 denoised_img = local_mean(noisy_img, size=size) 34 noise = (res**2).sum() / res.size
---> 31 l_var = local_var(noisy_img, size=size) ---> 35 noise_level = 1 - noise / l_var
(continues on next page) 36 noise_level[noise_level < 0] = 0
37 denoised_img = np.int64(noise_level * res)

9.3. Using the Python debugger 352 9.3. Using the Python debugger
FloatingPointError: divide by zero encountered in divide 353
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Getting help when in the debugger

Type h or help to access the interactive help:
Other ways of starting a debugger
ipdb> help
• Raising an exception as a poor man break point
If you find it tedious to note the line number to set a break point, you can simply raise an exception Documented commands (type help <topic>):
at the point that you want to inspect and use IPython’s %debug. Note that in this case you cannot ========================================
step or continue the execution. EOF commands enable ll pp s until
a condition exceptions longlist psource skip_hidden up
• Debugging test failures using nosetests alias cont exit n q skip_predicates w
You can run nosetests --pdb to drop in post-mortem debugging on exceptions, and nosetests args context h next quit source whatis
--pdb-failure to inspect test failures using the debugger. b continue help p r step where
break d ignore pdef restart tbreak
In addition, you can use the IPython interface for the debugger in nose by installing the nose plugin bt debug j pdoc return u
ipdbplugin. You can than pass --ipdb and --ipdb-failure options to nosetests. c disable jump pfile retval unalias
• Calling the debugger explicitly cl display l pinfo run undisplay
clear down list pinfo2 rv unt
Insert the following line where you want to drop in the debugger:

import pdb; pdb.set_trace() Miscellaneous help topics:

==========================
exec pdb

. Warning Undocumented commands:

When running nosetests, the output is captured, and thus it seems that the debugger does not ======================
work. Simply run the nosetests with the -s flag. interact

Graphical debuggers and alternatives

9.4 Debugging segmentation faults using gdb
• pudb is a good semi-graphical debugger with a text user interface in the console. If you have a segmentation fault, you cannot debug it with pdb, as it crashes the Python interpreter
before it can drop in the debugger. Similarly, if you have a bug in C code embedded in Python, pdb is
• The Visual Studio Code integrated development environment includes a debugging mode. useless. For this we turn to the gnu debugger, gdb, available on Linux.
• The Mu editor is a simple Python editor that includes a debugging mode. Before we start with gdb, let us add a few Python-specific tools to it. For this we add a few macros to
our ~/.gdbinit. The optimal choice of macro depends on your Python version and your gdb version. I
have added a simplified version in gdbinit, but feel free to read DebuggingWithGdb.
9.3.2 Debugger commands and interaction
To debug with gdb the Python script segfault.py, we can run the script in gdb as follows

l(list) Lists the code at the current position $ gdb python

u(p) Walk up the call stack ...
d(own) Walk down the call stack (gdb) run segfault.py
n(ext) Execute the next line (does not go down in new functions) Starting program: /usr/bin/python segfault.py
s(tep) Execute the next statement (goes down in new functions) [Thread debugging using libthread_db enabled]
bt Print the call stack
a Print the local variables Program received signal SIGSEGV, Segmentation fault.
!command Execute the given Python command (by opposition to pdb commands _strided_byte_copy (dst=0x8537478 "\360\343G", outstrides=4, src=
0x86c0690 <Address 0x86c0690 out of bounds>, instrides=32, N=3,
elsize=4)
at numpy/core/src/multiarray/ctors.c:365
. Warning 365 _FAST_MOVE(Int32);
(gdb)
Debugger commands are not Python code
You cannot name the variables the way you want. For instance, if in you cannot override the variables We get a segfault, and gdb captures it for post-mortem debugging in the C level stack (not the Python
in the current frame with the same name: use different names than your local variable when call stack). We can debug the C call stack using gdb’s commands:
typing code in the debugger.
(gdb) up
#1 0x004af4f5 in _copy_from_same_shape (dest=<value optimized out>,
(continues on next page)

9.3. Using the Python debugger 354 9.4. Debugging segmentation faults using gdb 355
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) Thus the segfault happens when printing big_array[-10:]. The reason is simply that big_array has
src=<value optimized out>, myfunc=0x496780 <_strided_byte_copy>, been allocated with its end outside the program memory.
swap=0)
at numpy/core/src/multiarray/ctors.c:748
ò Note
748 myfunc(dit->dataptr, dest->strides[maxaxis],
For a list of Python-specific commands defined in the gdbinit, read the source of this file.
As you can see, right now, we are in the C code of numpy. We would like to know what is the Python
code that triggers this segfault, so we go up the stack until we hit the Python execution loop:

(gdb) up
#8 0x080ddd23 in call_function (f= Wrap up exercise
Frame 0x85371ec, for file /home/varoquau/usr/lib/python2.6/site-packages/numpy/
˓→core/arrayprint.py, line 156, in _leading_trailing (a=<numpy.ndarray at remote␣ The following script is well documented and hopefully legible. It seeks to answer a problem of actual
˓→0x85371b0>, _nc=<module at remote 0xb7f93a64>), throwflag=0) interest for numerical computing, but it does not work. . . Can you debug it?
at ../Python/ceval.c:3750 Python source code: to_debug.py
3750 ../Python/ceval.c: No such file or directory.
in ../Python/ceval.c

(gdb) up
#9 PyEval_EvalFrameEx (f=
Frame 0x85371ec, for file /home/varoquau/usr/lib/python2.6/site-packages/numpy/
˓→core/arrayprint.py, line 156, in _leading_trailing (a=<numpy.ndarray at remote␣

˓→0x85371b0>, _nc=<module at remote 0xb7f93a64>), throwflag=0)

at ../Python/ceval.c:2412
2412 in ../Python/ceval.c
(gdb)

Once we are in the Python execution loop, we can use our special Python helper function. For instance
we can find the corresponding Python code:

(gdb) pyframe
/home/varoquau/usr/lib/python2.6/site-packages/numpy/core/arrayprint.py (158): _
˓→leading_trailing

(gdb)

This is numpy code, we need to go up until we find code that we have written:

(gdb) up
...
(gdb) up
#34 0x080dc97a in PyEval_EvalFrameEx (f=
Frame 0x82f064c, for file segfault.py, line 11, in print_big_array (small_array=
˓→<numpy.ndarray at remote 0x853ecf0>, big_array=<numpy.ndarray at remote 0x853ed20>),

˓→ throwflag=0) at ../Python/ceval.c:1630

1630 ../Python/ceval.c: No such file or directory.

in ../Python/ceval.c
(gdb) pyframe
segfault.py (12): print_big_array

The corresponding code is:

def make_big_array(small_array):
big_array = stride_tricks.as_strided(
small_array, shape=(int(2e6), int(2e6)), strides=(32, 32)
)
return big_array

9.4. Debugging segmentation faults using gdb 356 9.4. Debugging segmentation faults using gdb 357
 Scientific Python Lectures, Edition 2025.1rc0.dev0

• Writing faster numerical code

– Additional Links

10
10.1 Optimization workflow
1. Make it work: write the code in a simple legible ways.
2. Make it work reliably: write automated test cases, make really sure that your algorithm is right
and that if you break it, the tests will capture the breakage.
CHAPTER
3. Optimize the code by profiling simple use-cases to find the bottlenecks and speeding up these
bottleneck, finding a better algorithm or implementation. Keep in mind that a trade off should
be found between profiling on a realistic example and the simplicity and speed of execution of the
code. For efficient work, it is best to work with profiling runs lasting around 10s.

10.2 Profiling Python code

Optimizing code No optimization without measuring!

• Measure: profiling, timing

• You’ll have surprises: the fastest code is not always what you think

10.2.1 Timeit
In IPython, use timeit (https://docs.python.org/3/library/timeit.html) to time elementary operations:
Donald Knuth In [1]: import numpy as np
“Premature optimization is the root of all evil”
In [2]: a = np.arange(1000)
Author: Gaël Varoquaux In [3]: %timeit a ** 2
This chapter deals with strategies to make Python code go faster. 950 ns +- 1.49 ns per loop (mean +- std. dev. of 7 runs, 1,000,000 loops each)

In [4]: %timeit a ** 2.1

Prerequisites 15.9 us +- 27.2 ns per loop (mean +- std. dev. of 7 runs, 100,000 loops each)
• line_profiler
In [5]: %timeit a * a
1.04 us +- 2.29 ns per loop (mean +- std. dev. of 7 runs, 1,000,000 loops each)

Chapters contents Use this to guide your choice between strategies.

• Optimization workflow
ò Note
• Profiling Python code
For long running calls, using %time instead of %timeit; it is less precise but faster
– Timeit
– Profiler
– Line-profiler
10.2.2 Profiler
• Making code go faster Useful when you have a large program to profile, for example the following file:

– Algorithmic optimization # For this example to run, you also need the 'ica.py' file
∗ Example of the SVD import numpy as np
(continues on next page)

358 10.1. Optimization workflow 359

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

import scipy as sp 1 0.000 0.000 14.551 14.551 demo.py:1 (<module>)
29 0.000 0.000 0.000 0.000 numeric.py:180 (asarray)
from ica import fastica 35 0.000 0.000 0.000 0.000 defmatrix.py:193 (__new__)
35 0.000 0.000 0.001 0.000 defmatrix.py:43 (asmatrix)
21 0.000 0.000 0.001 0.000 defmatrix.py:287 (__mul__)
# @profile # uncomment this line to run with line_profiler 41 0.000 0.000 0.000 0.000 {numpy.core.multiarray.zeros}
def test(): 28 0.000 0.000 0.000 0.000 {method 'transpose' of 'numpy.ndarray'␣
rng = np.random.default_rng() ˓→objects}

data = rng.random((5000, 100)) 1 0.000 0.000 0.008 0.008 ica.py:97 (fastica)

u, s, v = sp.linalg.svd(data) ...
pca = u[:, :10].T @ data
results = fastica(pca.T, whiten=False) Clearly the svd (in decomp.py) is what takes most of our time, a.k.a. the bottleneck. We have to find a
way to make this step go faster, or to avoid this step (algorithmic optimization). Spending time on the
rest of the code is useless.
if __name__ == "__main__":
test()
Profiling outside of IPython, running ``cProfile``

Similar profiling can be done outside of IPython, simply calling the built-in Python profilers cProfile
ò Note and profile.
This is a combination of two unsupervised learning techniques, principal component analysis (PCA) $ python -m cProfile -o demo.prof demo.py
and independent component analysis (ICA). PCA is a technique for dimensionality reduction, i.e.
an algorithm to explain the observed variance in your data using less dimensions. ICA is a source Using the -o switch will output the profiler results to the file demo.prof to view with an external
separation technique, for example to unmix multiple signals that have been recorded through multiple tool. This can be useful if you wish to process the profiler output with a visualization tool.
sensors. Doing a PCA first and then an ICA can be useful if you have more sensors than signals. For
more information see: the FastICA example from scikits-learn.
10.2.3 Line-profiler
To run it, you also need to download the ica module. In IPython we can time the script: The profiler tells us which function takes most of the time, but not where it is called.
For this, we use the line_profiler: in the source file, we decorate a few functions that we want to inspect
In [6]: %run -t demo.py with @profile (no need to import it)
IPython CPU timings (estimated):
User : 14.3929 s. @profile
System: 0.256016 s. def test():
rng = np.random.default_rng()
and profile it: data = rng.random((5000, 100))
u, s, v = linalg.svd(data)
In [7]: %run -p demo.py pca = u[:, :10] @ data
916 function calls in 14.551 CPU seconds results = fastica(pca.T, whiten=False)
Ordered by: internal time
ncalls tottime percall cumtime percall filename:lineno (function) Then we run the script using the kernprof command, with switches -l, --line-by-line and -v,
1 14.457 14.457 14.479 14.479 decomp.py:849 (svd) --view to use the line-by-line profiler and view the results in addition to saving them:
1 0.054 0.054 0.054 0.054 {method 'random_sample' of 'mtrand.
˓→RandomState' objects} $ kernprof -l -v demo.py
1 0.017 0.017 0.021 0.021 function_base.py:645 (asarray_chkfinite)
54 0.011 0.000 0.011 0.000 {numpy.core._dotblas.dot} Wrote profile results to demo.py.lprof
2 0.005 0.002 0.005 0.002 {method 'any' of 'numpy.ndarray' objects} Timer unit: 1e-06 s
6 0.001 0.000 0.001 0.000 ica.py:195 (gprime)
6 0.001 0.000 0.001 0.000 ica.py:192 (g) Total time: 1.27874 s
14 0.001 0.000 0.001 0.000 {numpy.linalg.lapack_lite.dsyevd} File: demo.py
19 0.001 0.000 0.001 0.000 twodim_base.py:204 (diag) Function: test at line 9
1 0.001 0.001 0.008 0.008 ica.py:69 (_ica_par)
1 0.001 0.001 14.551 14.551 {execfile} Line # Hits Time Per Hit % Time Line Contents
107 0.000 0.000 0.001 0.000 defmatrix.py:239 (__array_finalize__) ==============================================================
7 0.000 0.000 0.004 0.001 ica.py:58 (_sym_decorrelation) 9 @profile
7 0.000 0.000 0.002 0.000 linalg.py:841 (eigh) 10 def test():
172 0.000 0.000 0.000 0.000 {isinstance} 11 1 69.0 69.0 0.0 rng = np.random.default_rng()
(continues on next page) (continues on next page)

10.2. Profiling Python code 360 10.2. Profiling Python code 361
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

12 1 2453.0 2453.0 0.2 data = rng.random((5000, 100)) demo.linalg demo.prof demo.prof.png demo.py.lprof demo.test
13 1 1274715.0 1274715.0 99.7 u, s, v = sp.linalg.svd(data)
14 1 413.0 413.0 0.0 pca = u[:, :10].T @ data In [15]: %timeit demo.test()
15 1 1094.0 1094.0 0.1 results = fastica(pca.T,␣ ica.py:65: RuntimeWarning: invalid value encountered in sqrt
˓→whiten=False) W = (u * np.diag(1.0/np.sqrt(s)) * u.T) * W # W = (W * W.T) ^{-1/2} * W
1 loops, best of 3: 17.5 s per loop
The SVD is taking all the time. We need to optimise this line.
In [16]: import demo_opt

10.3 Making code go faster In [17]: %timeit demo_opt.test()

1 loops, best of 3: 208 ms per loop
Once we have identified the bottlenecks, we need to make the corresponding code go faster.
Real incomplete SVDs, e.g. computing only the first 10 eigenvectors, can be computed with arpack,
10.3.1 Algorithmic optimization available in scipy.sparse.linalg.eigsh.
The first thing to look for is algorithmic optimization: are there ways to compute less, or better?
For a high-level view of the problem, a good understanding of the maths behind the algorithm helps. Computational linear algebra
However, it is not uncommon to find simple changes, like moving computation or memory allocation For certain algorithms, many of the bottlenecks will be linear algebra computations. In this case,
outside a for loop, that bring in big gains. using the right function to solve the right problem is key. For instance, an eigenvalue problem with
a symmetric matrix is easier to solve than with a general matrix. Also, most often, you can avoid
Example of the SVD inverting a matrix and use a less costly (and more numerically stable) operation.
In both examples above, the SVD - Singular Value Decomposition - is what takes most of the time. Know your computational linear algebra. When in doubt, explore scipy.linalg, and use %timeit
Indeed, the computational cost of this algorithm is roughly 𝑛3 in the size of the input matrix. to try out different alternatives on your data.
However, in both of these example, we are not using all the output of the SVD, but only the first few rows
of its first return argument. If we use the svd implementation of SciPy, we can ask for an incomplete
version of the SVD. Note that implementations of linear algebra in SciPy are richer then those in NumPy 10.4 Writing faster numerical code
and should be preferred.
A complete discussion on advanced use of NumPy is found in chapter Advanced NumPy, or in the article
In [8]: %timeit np.linalg.svd(data) The NumPy array: a structure for efficient numerical computation by van der Walt et al. Here we discuss
1 loops, best of 3: 14.5 s per loop only some commonly encountered tricks to make code faster.

In [9]: import scipy as sp • Vectorizing for loops

Find tricks to avoid for loops using NumPy arrays. For this, masks and indices arrays can be
In [10]: %timeit sp.linalg.svd(data) useful.
1 loops, best of 3: 14.2 s per loop
• Broadcasting
In [11]: %timeit sp.linalg.svd(data, full_matrices=False) Use broadcasting to do operations on arrays as small as possible before combining them.
1 loops, best of 3: 295 ms per loop
• In place operations
In [12]: %timeit np.linalg.svd(data, full_matrices=False)
In [18]: a = np.zeros(1e7)
1 loops, best of 3: 293 ms per loop
In [19]: %timeit global a ; a = 0*a
We can then use this insight to optimize the previous code: 10 loops, best of 3: 111 ms per loop
def test():
rng = np.random.default_rng() In [20]: %timeit global a ; a *= 0
data = rng.random((5000, 100)) 10 loops, best of 3: 48.4 ms per loop
u, s, v = sp.linalg.svd(data, full_matrices=False)
pca = u[:, :10].T @ data note: we need global a in the timeit so that it work, as it is assigning to a, and thus considers it
results = fastica(pca.T, whiten=False) as a local variable.
• Be easy on the memory: use views, and not copies
In [13]: import demo Copying big arrays is as costly as making simple numerical operations on them:
In [14]: %timeit demo. In [21]: a = np.zeros(1e7)
demo.fastica demo.np demo.prof.pdf demo.py demo.pyc
(continues on next page) (continues on next page)

10.3. Making code go faster 362 10.4. Writing faster numerical code 363
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) 10.4.1 Additional Links

In [22]: %timeit a.copy() • If you need to profile memory usage, you could try the memory_profiler
10 loops, best of 3: 124 ms per loop
• If you need to profile down into C extensions, you could try using gperftools from Python with
In [23]: %timeit a + 1 yep.
10 loops, best of 3: 112 ms per loop • If you would like to track performance of your code across time, i.e. as you make new commits to
your repository, you could try: asv
• Beware of cache effects
• If you need some interactive visualization why not try RunSnakeRun
Memory access is cheaper when it is grouped: accessing a big array in a continuous way is much
faster than random access. This implies amongst other things that smaller strides are faster
(see CPU cache effects):

In [24]: c = np.zeros((1e4, 1e4), order='C')

In [25]: %timeit c.sum(axis=0)

1 loops, best of 3: 3.89 s per loop

In [26]: %timeit c.sum(axis=1)

1 loops, best of 3: 188 ms per loop

In [27]: c.strides
Out[27]: (80000, 8)

This is the reason why Fortran ordering or C ordering may make a big difference on operations:

In [28]: rng = np.random.default_rng()

In [29]: a = rng.random((20, 2**18))

In [30]: b = rng.random((20, 2**18))

In [31]: %timeit b @ a.T

9.32 ms +- 49.6 us per loop (mean +- std. dev. of 7 runs, 100 loops each)

In [32]: c = np.ascontiguousarray(a.T)

In [33]: %timeit b @ c
10.1 ms +- 158 us per loop (mean +- std. dev. of 7 runs, 100 loops each)

Note that copying the data to work around this effect may not be worth it:

In [34]: %timeit c = np.ascontiguousarray(a.T)

14.4 ms +- 24.9 us per loop (mean +- std. dev. of 7 runs, 100 loops each)

Using numexpr can be useful to automatically optimize code for such effects.
• Use compiled code
The last resort, once you are sure that all the high-level optimizations have been explored, is to
transfer the hot spots, i.e. the few lines or functions in which most of the time is spent, to compiled
code. For compiled code, the preferred option is to use Cython: it is easy to transform exiting
Python code in compiled code, and with a good use of the NumPy support yields efficient code on
NumPy arrays, for instance by unrolling loops.

. Warning

For all the above: profile and time your choices. Don’t base your optimization on theoretical consid-
erations.

10.4. Writing faster numerical code 364 10.4. Writing faster numerical code 365
 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

[<matplotlib.lines.Line2D object at ...>]
>>> plt.xlabel('size n')
Text(...'size n')
>>> plt.ylabel('memory [MB]')

11
Text(...'memory [MB]')

11.1.2 Sparse Matrices vs. Sparse Matrix Storage Schemes

• sparse matrix is a matrix, which is almost empty

CHAPTER • storing all the zeros is wasteful -> store only nonzero items
• think compression
• pros: huge memory savings
• cons: slow access to individual items, but it depends on actual storage scheme.

11.1.3 Typical Applications

• solution of partial differential equations (PDEs)
Sparse Arrays in SciPy – the finite element method
– mechanical engineering, electrotechnics, physics, . . .
• graph theory
– nonzero at (i, j) means that node i is connected to node j
• natural language processing

Author: Robert Cimrman – nonzero at (i, j) means that the document i contains the word j
• ...

11.1.4 Prerequisites
• numpy
11.1 Introduction • scipy

(dense) matrix is: • matplotlib (optional)

• mathematical object • ipython (the enhancements come handy)

• data structure for storing a 2D array of values

11.1.5 Sparsity Structure Visualization
important features:
• spy() from matplotlib
• memory allocated once for all items
• example plots:
– usually a contiguous chunk, think NumPy ndarray
• fast access to individual items (*)

11.1.1 Why Sparse Matrices?

• the memory grows like n**2 for dense matrix
• small example (double precision matrix):
>>> import numpy as np
>>> import matplotlib.pyplot as plt
>>> x = np.linspace(0, 1e6, 10)
>>> plt.plot(x, 8.0 * (x**2) / 1e6, lw=5)
(continues on next page)

366 11.1. Introduction 367

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

7. dia_array: DIAgonal format

• each suitable for some tasks
• many employ sparsetools C++ module by Nathan Bell
• assume the following is imported:

>>> import numpy as np

>>> import scipy as sp
>>> import matplotlib.pyplot as plt

• warning for Numpy users:

– passing a sparse array object to NumPy functions that expect ndarray/matrix does not
work. Use sparse functions.
– the older csr_matrix classes use ‘*’ for matrix multiplication (dot product) and
‘A.multiply(B)’ for elementwise multiplication.
– the newer csr_array uses ‘@’ for dot product and ‘*’ for elementwise multiplication
– sparse arrays can be 1D or 2D, but not nD for n > 2 (unlike Numpy arrays).

11.2.1 Common Methods

• all scipy.sparse array classes are subclasses of sparray
– default implementation of arithmetic operations
∗ always converts to CSR
∗ subclasses override for efficiency
– shape, data type, set/get
– indices of nonzero values in the array
– format conversion, interaction with NumPy (toarray())
– ...
• attributes:
– mtx.T - transpose (same as mtx.transpose())
– mtx.real - real part of complex matrix
– mtx.imag - imaginary part of complex matrix
– mtx.size - the number of nonzeros (same as self.getnnz())
– mtx.shape - the number of rows and columns (tuple)
• data and indices usually stored in 1D NumPy arrays

11.2 Storage Schemes 11.2.2 Sparse Array Classes

• seven sparse array types in scipy.sparse: Diagonal Format (DIA)
1. csr_array: Compressed Sparse Row format • very simple scheme
2. csc_array: Compressed Sparse Column format • diagonals in dense NumPy array of shape (n_diag, length)
3. bsr_array: Block Sparse Row format – fixed length -> waste space a bit when far from main diagonal
4. lil_array: List of Lists format – subclass of _data_matrix (sparse array classes with .data attribute)
5. dok_array: Dictionary of Keys format • offset for each diagonal
6. coo_array: COOrdinate format (aka IJV, triplet format) – 0 is the main diagonal

11.2. Storage Schemes 368 11.2. Storage Schemes 369

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

– negative offset = below (continued from previous page)

– positive offset = above with 9 stored elements (3 diagonals) and shape (4, 4)>
Coords Values
• fast matrix * vector (sparsetools) (0, 0) 1
• fast and easy item-wise operations (1, 1) 2
(2, 2) 3
– manipulate data array directly (fast NumPy machinery) (3, 3) 4
• constructor accepts: (1, 0) 5
(2, 1) 6
– dense array/matrix (3, 2) 7
– sparse array/matrix (0, 2) 11
(1, 3) 12
– shape tuple (create empty array) >>> mtx.toarray()
– (data, offsets) tuple array([[ 1, 0, 11, 0],
[ 5, 2, 0, 12],
• no slicing, no individual item access [ 0, 6, 3, 0],
• use: [ 0, 0, 7, 4]])

– rather specialized • explanation with a scheme:

– solving PDEs by finite differences
offset: row
– with an iterative solver
2: 9
Examples 1: --10------
0: 1 . 11 .
• create some DIA arrays: -1: 5 2 . 12
-2: . 6 3 .
>>> data = np.array([[1, 2, 3, 4]]).repeat(3, axis=0) -3: . . 7 4
>>> data ---------8
array([[1, 2, 3, 4],
[1, 2, 3, 4], • matrix-vector multiplication
[1, 2, 3, 4]])
>>> offsets = np.array([0, -1, 2]) >>> vec = np.ones((4, ))
>>> mtx = sp.sparse.dia_array((data, offsets), shape=(4, 4)) >>> vec
>>> mtx array([1., 1., 1., 1.])
<DIAgonal sparse array of dtype 'int64' >>> mtx @ vec
with 9 stored elements (3 diagonals) and shape (4, 4)> array([12., 19., 9., 11.])
>>> mtx.toarray() >>> (mtx * vec).toarray()
array([[1, 0, 3, 0], array([[ 1., 0., 11., 0.],
[1, 2, 0, 4], [ 5., 2., 0., 12.],
[0, 2, 3, 0], [ 0., 6., 3., 0.],
[0, 0, 3, 4]]) [ 0., 0., 7., 4.]])

>>> data = np.arange(12).reshape((3, 4)) + 1

>>> data
array([[ 1, 2, 3, 4], List of Lists Format (LIL)
[ 5, 6, 7, 8],
[ 9, 10, 11, 12]]) • row-based linked list
>>> mtx = sp.sparse.dia_array((data, offsets), shape=(4, 4)) – each row is a Python list (sorted) of column indices of non-zero elements
>>> mtx.data
array([[ 1, 2, 3, 4], – rows stored in a NumPy array (dtype=np.object)
[ 5, 6, 7, 8], – non-zero values data stored analogously
[ 9, 10, 11, 12]])
>>> mtx.offsets • efficient for constructing sparse arrays incrementally
array([ 0, -1, 2], dtype=int32) • constructor accepts:
>>> print(mtx)
<DIAgonal sparse array of dtype 'int64' – dense array/matrix
(continues on next page)
– sparse array/matrix

11.2. Storage Schemes 370 11.2. Storage Schemes 371

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

– shape tuple (create empty array) (continued from previous page)

• flexible slicing, changing sparsity structure is efficient (0, 1) 1

(0, 2) 2
• slow arithmetic, slow column slicing due to being row-based (1, 0) 3
• use: (1, 2) 1
(2, 0) 1
– when sparsity pattern is not known apriori or changes (2, 3) 1
– example: reading a sparse array from a text file >>> mtx[:2, :]
<List of Lists sparse array of dtype 'int64'
Examples with 4 stored elements and shape (2, 4)>
>>> mtx[:2, :].toarray()
• create an empty LIL array: array([[0, 1, 2, 0],
[3, 0, 1, 0]]...)
>>> mtx = sp.sparse.lil_array((4, 5)) >>> mtx[1:2, [0,2]].toarray()
array([[3, 1]]...)
• prepare random data: >>> mtx.toarray()
array([[0, 1, 2, 0],
>>> rng = np.random.default_rng(27446968) [3, 0, 1, 0],
>>> data = np.round(rng.random((2, 3))) [1, 0, 0, 1]]...)
>>> data
array([[1., 0., 1.],
[0., 0., 1.]])
Dictionary of Keys Format (DOK)
• assign the data using fancy indexing:
• subclass of Python dict
>>> mtx[:2, [1, 2, 3]] = data
>>> mtx – keys are (row, column) index tuples (no duplicate entries allowed)
<List of Lists sparse array of dtype 'float64' – values are corresponding non-zero values
with 3 stored elements and shape (4, 5)>
>>> print(mtx) • efficient for constructing sparse arrays incrementally
<List of Lists sparse array of dtype 'float64' • constructor accepts:
with 3 stored elements and shape (4, 5)>
Coords Values – dense array/matrix
(0, 1) 1.0 – sparse array/matrix
(0, 3) 1.0
(1, 3) 1.0 – shape tuple (create empty array)
>>> mtx.toarray() • efficient O(1) access to individual elements
array([[0., 1., 0., 1., 0.],
[0., 0., 0., 1., 0.], • flexible slicing, changing sparsity structure is efficient
[0., 0., 0., 0., 0.], • can be efficiently converted to a coo_array once constructed
[0., 0., 0., 0., 0.]])
>>> mtx.toarray() • slow arithmetic (for loops with dict.items())
array([[0., 1., 0., 1., 0.], • use:
[0., 0., 0., 1., 0.],
[0., 0., 0., 0., 0.], – when sparsity pattern is not known apriori or changes
[0., 0., 0., 0., 0.]])
Examples
• more slicing and indexing:
• create a DOK array element by element:
>>> mtx = sp.sparse.lil_array([[0, 1, 2, 0], [3, 0, 1, 0], [1, 0, 0, 1]])
>>> mtx.toarray() >>> mtx = sp.sparse.dok_array((5, 5), dtype=np.float64)
array([[0, 1, 2, 0], >>> mtx
[3, 0, 1, 0], <Dictionary Of Keys sparse array of dtype 'float64'
[1, 0, 0, 1]]...) with 0 stored elements and shape (5, 5)>
>>> print(mtx) >>> for ir in range(5):
<List of Lists sparse array of dtype 'int64' ... for ic in range(5):
with 6 stored elements and shape (3, 4)> ... mtx[ir, ic] = 1.0 * (ir != ic)
Coords Values >>> mtx
(continues on next page) (continues on next page)

11.2. Storage Schemes 372 11.2. Storage Schemes 373

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) ∗ facilitates efficient construction of finite element matrices
<Dictionary Of Keys sparse array of dtype 'float64'
with 20 stored elements and shape (5, 5)> Examples
>>> mtx.toarray()
array([[0., 1., 1., 1., 1.], • create empty COO array:
[1., 0., 1., 1., 1.],
>>> mtx = sp.sparse.coo_array((3, 4), dtype=np.int8)
[1., 1., 0., 1., 1.],
>>> mtx.toarray()
[1., 1., 1., 0., 1.],
array([[0, 0, 0, 0],
[1., 1., 1., 1., 0.]])
[0, 0, 0, 0],
[0, 0, 0, 0]], dtype=int8)
• slicing and indexing:

>>> mtx[1, 1] • create using (data, ij) tuple:

np.float64(0.0)
>>> row = np.array([0, 3, 1, 0])
>>> mtx[[1], 1:3]
>>> col = np.array([0, 3, 1, 2])
<Dictionary Of Keys sparse array of dtype 'float64'
>>> data = np.array([4, 5, 7, 9])
with 1 stored elements and shape (1, 2)>
>>> mtx = sp.sparse.coo_array((data, (row, col)), shape=(4, 4))
>>> mtx[[1], 1:3].toarray()
>>> mtx
array([[0., 1.]])
<COOrdinate sparse array of dtype 'int64'
>>> mtx[[2, 1], 1:3].toarray()
with 4 stored elements and shape (4, 4)>
array([[1., 0.],
>>> mtx.toarray()
[0., 1.]])
array([[4, 0, 9, 0],
[0, 7, 0, 0],
[0, 0, 0, 0],
[0, 0, 0, 5]])
Coordinate Format (COO)
• also known as the ‘ijv’ or ‘triplet’ format • duplicates entries are summed together:

– three NumPy arrays: row, col, data. >>> row = np.array([0, 0, 1, 3, 1, 0, 0])
>>> col = np.array([0, 2, 1, 3, 1, 0, 0])
– attribute coords is the tuple (row, col)
>>> data = np.array([1, 1, 1, 1, 1, 1, 1])
– data[i] is value at (row[i], col[i]) position >>> mtx = sp.sparse.coo_array((data, (row, col)), shape=(4, 4))
>>> mtx.toarray()
– permits duplicate entries
array([[3, 0, 1, 0],
– subclass of _data_matrix (sparse matrix classes with .data attribute) [0, 2, 0, 0],
[0, 0, 0, 0],
• fast format for constructing sparse arrays
[0, 0, 0, 1]])
• constructor accepts:
– dense array/matrix • no slicing. . . :

– sparse array/matrix >>> mtx[2, 3]

Traceback (most recent call last):
– shape tuple (create empty matrix) ...
– (data, coords) tuple TypeError: 'coo_array' object ...
• very fast conversion to and from CSR/CSC formats
• fast matrix * vector (sparsetools)
• fast and easy item-wise operations Compressed Sparse Row Format (CSR)

– manipulate data array directly (fast NumPy machinery) • row oriented

• no slicing, no arithmetic (directly, converts to CSR) – three NumPy arrays: indices, indptr, data

• use: ∗ indices is array of column indices

– facilitates fast conversion among sparse formats ∗ data is array of corresponding nonzero values

– when converting to other format (usually CSR or CSC), duplicate entries are summed ∗ indptr points to row starts in indices and data
together ∗ length of indptr is n_row + 1, last item = number of values = length of both
indices and data

11.2. Storage Schemes 374 11.2. Storage Schemes 375

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

∗ nonzero values of the i-th row are data[indptr[i]:indptr[i + 1]] with column indices
>>> data = np.array([1, 2, 3, 4, 5, 6])
indices[indptr[i]:indptr[i + 1]]
>>> indices = np.array([0, 2, 2, 0, 1, 2])
∗ item (i, j) can be accessed as data[indptr[i] + k], where k is position of j in in- >>> indptr = np.array([0, 2, 3, 6])
dices[indptr[i]:indptr[i + 1]] >>> mtx = sp.sparse.csr_array((data, indices, indptr), shape=(3, 3))
>>> mtx.toarray()
– subclass of _cs_matrix (common CSR/CSC functionality)
array([[1, 0, 2],
∗ subclass of _data_matrix (sparse array classes with .data attribute) [0, 0, 3],
[4, 5, 6]])
• fast matrix vector products and other arithmetic (sparsetools)
• constructor accepts:
– dense array/matrix
Compressed Sparse Column Format (CSC)
– sparse array/matrix
• column oriented
– shape tuple (create empty array)
– three NumPy arrays: indices, indptr, data
– (data, coords) tuple
∗ indices is array of row indices
– (data, indices, indptr) tuple
∗ data is array of corresponding nonzero values
• efficient row slicing, row-oriented operations
∗ indptr points to column starts in indices and data
• slow column slicing, expensive changes to the sparsity structure
∗ length is n_col + 1, last item = number of values = length of both indices and
• use: data
– actual computations (most linear solvers support this format) ∗ nonzero values of the i-th column are data[indptr[i]:indptr[i+1]] with row indices
indices[indptr[i]:indptr[i+1]]
Examples
∗ item (i, j) can be accessed as data[indptr[j]+k], where k is position of i in in-
• create empty CSR array: dices[indptr[j]:indptr[j+1]]
– subclass of _cs_matrix (common CSR/CSC functionality)
>>> mtx = sp.sparse.csr_array((3, 4), dtype=np.int8)
>>> mtx.toarray() ∗ subclass of _data_matrix (sparse array classes with .data attribute)
array([[0, 0, 0, 0],
• fast matrix vector products and other arithmetic (sparsetools)
[0, 0, 0, 0],
[0, 0, 0, 0]], dtype=int8) • constructor accepts:
– dense array/matrix
• create using (data, coords) tuple:
– sparse array/matrix
>>> row = np.array([0, 0, 1, 2, 2, 2])
>>> col = np.array([0, 2, 2, 0, 1, 2]) – shape tuple (create empty array)
>>> data = np.array([1, 2, 3, 4, 5, 6]) – (data, coords) tuple
>>> mtx = sp.sparse.csr_array((data, (row, col)), shape=(3, 3))
>>> mtx – (data, indices, indptr) tuple
<Compressed Sparse Row sparse array of dtype 'int64' • efficient column slicing, column-oriented operations
with 6 stored elements and shape (3, 3)>
>>> mtx.toarray() • slow row slicing, expensive changes to the sparsity structure
array([[1, 0, 2], • use:
[0, 0, 3],
[4, 5, 6]]...) – actual computations (most linear solvers support this format)
>>> mtx.data
array([1, 2, 3, 4, 5, 6]...) Examples
>>> mtx.indices
array([0, 2, 2, 0, 1, 2]) • create empty CSC array:
>>> mtx.indptr
>>> mtx = sp.sparse.csc_array((3, 4), dtype=np.int8)
array([0, 2, 3, 6])
>>> mtx.toarray()
array([[0, 0, 0, 0],
• create using (data, indices, indptr) tuple: [0, 0, 0, 0],
[0, 0, 0, 0]], dtype=int8)

11.2. Storage Schemes 376 11.2. Storage Schemes 377

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• create using (data, coords) tuple: • many arithmetic operations considerably more efficient than CSR for sparse matrices with dense
sub-matrices
>>> row = np.array([0, 0, 1, 2, 2, 2])
>>> col = np.array([0, 2, 2, 0, 1, 2]) • use:
>>> data = np.array([1, 2, 3, 4, 5, 6]) – like CSR
>>> mtx = sp.sparse.csc_array((data, (row, col)), shape=(3, 3))
>>> mtx – vector-valued finite element discretizations
<Compressed Sparse Column sparse array of dtype 'int64'
with 6 stored elements and shape (3, 3)> Examples
>>> mtx.toarray()
array([[1, 0, 2], • create empty BSR array with (1, 1) block size (like CSR. . . ):
[0, 0, 3],
>>> mtx = sp.sparse.bsr_array((3, 4), dtype=np.int8)
[4, 5, 6]]...)
>>> mtx
>>> mtx.data
<Block Sparse Row sparse array of dtype 'int8'
array([1, 4, 5, 2, 3, 6]...)
with 0 stored elements (blocksize=1x1) and shape (3, 4)>
>>> mtx.indices
>>> mtx.toarray()
array([0, 2, 2, 0, 1, 2])
array([[0, 0, 0, 0],
>>> mtx.indptr
[0, 0, 0, 0],
array([0, 2, 3, 6])
[0, 0, 0, 0]], dtype=int8)
• create using (data, indices, indptr) tuple:
• create empty BSR array with (3, 2) block size:
>>> data = np.array([1, 4, 5, 2, 3, 6])
>>> mtx = sp.sparse.bsr_array((3, 4), blocksize=(3, 2), dtype=np.int8)
>>> indices = np.array([0, 2, 2, 0, 1, 2])
>>> mtx
>>> indptr = np.array([0, 2, 3, 6])
<Block Sparse Row sparse array of dtype 'int8'
>>> mtx = sp.sparse.csc_array((data, indices, indptr), shape=(3, 3))
with 0 stored elements (blocksize=3x2) and shape (3, 4)>
>>> mtx.toarray()
>>> mtx.toarray()
array([[1, 0, 2],
array([[0, 0, 0, 0],
[0, 0, 3],
[0, 0, 0, 0],
[4, 5, 6]])
[0, 0, 0, 0]], dtype=int8)

– a bug?

Block Compressed Row Format (BSR) • create using (data, coords) tuple with (1, 1) block size (like CSR. . . ):

• basically a CSR with dense sub-matrices of fixed shape instead of scalar items >>> row = np.array([0, 0, 1, 2, 2, 2])
>>> col = np.array([0, 2, 2, 0, 1, 2])
– block size (R, C) must evenly divide the shape of the matrix (M, N)
>>> data = np.array([1, 2, 3, 4, 5, 6])
– three NumPy arrays: indices, indptr, data >>> mtx = sp.sparse.bsr_array((data, (row, col)), shape=(3, 3))
>>> mtx
∗ indices is array of column indices for each block
<Block Sparse Row sparse array of dtype 'int64'
∗ data is array of corresponding nonzero values of shape (nnz, R, C) with 6 stored elements (blocksize=1x1) and shape (3, 3)>
>>> mtx.toarray()
∗ ...
array([[1, 0, 2],
– subclass of _cs_matrix (common CSR/CSC functionality) [0, 0, 3],
[4, 5, 6]]...)
∗ subclass of _data_matrix (sparse matrix classes with .data attribute)
>>> mtx.data
• fast matrix vector products and other arithmetic (sparsetools) array([[[1]],
• constructor accepts:
[[2]],
– dense array/matrix
[[3]],
– sparse array/matrix
– shape tuple (create empty array) [[4]],
– (data, coords) tuple
[[5]],
– (data, indices, indptr) tuple
[[6]]]...)
(continues on next page)

11.2. Storage Schemes 378 11.2. Storage Schemes 379

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) 11.2.3 Summary

>>> mtx.indices
array([0, 2, 2, 0, 1, 2]) Table 1: Summary of storage schemes.
>>> mtx.indptr
format matrix get fancy set item fancy solvers note
array([0, 2, 3, 6])
* vector item get set
• create using (data, indices, indptr) tuple with (2, 2) block size: CSR sparse- yes yes slow . any has data array, fast row-wise
tools ops
>>> indptr = np.array([0, 2, 3, 6]) CSC sparse- yes yes slow . any has data array, fast column-
>>> indices = np.array([0, 2, 2, 0, 1, 2]) tools wise ops
>>> data = np.array([1, 2, 3, 4, 5, 6]).repeat(4).reshape(6, 2, 2) BSR sparse- . . . . special- has data array, specialized
>>> mtx = sp.sparse.bsr_array((data, indices, indptr), shape=(6, 6)) tools ized
>>> mtx.toarray() COO sparse- . . . . itera- has data array, facilitates
array([[1, 1, 0, 0, 2, 2], tools tive fast conversion
[1, 1, 0, 0, 2, 2], DIA sparse- . . . . itera- has data array, specialized
[0, 0, 0, 0, 3, 3], tools tive
[0, 0, 0, 0, 3, 3], LIL via yes yes yes yes itera- arithmetic via CSR, incre-
[4, 4, 5, 5, 6, 6], CSR tive mental construction
[4, 4, 5, 5, 6, 6]]) DOK python yes one yes yes itera- O(1) item access, incremen-
>>> data axis tive tal construction, slow arith-
array([[[1, 1], only metic
[1, 1]],

[[2, 2],
[2, 2]],

[[3, 3],
11.3 Linear System Solvers
[3, 3]], • sparse matrix/eigenvalue problem solvers live in scipy.sparse.linalg
[[4, 4], • the submodules:
[4, 4]], – dsolve: direct factorization methods for solving linear systems
[[5, 5], – isolve: iterative methods for solving linear systems
[5, 5]], – eigen: sparse eigenvalue problem solvers
[[6, 6], • all solvers are accessible from:
[6, 6]]])
>>> import scipy as sp
>>> sp.sparse.linalg.__all__
['ArpackError', 'ArpackNoConvergence', ..., 'use_solver']

11.3.1 Sparse Direct Solvers

• default solver: SuperLU
– included in SciPy
– real and complex systems
– both single and double precision
• optional: umfpack
– real and complex systems
– double precision only
– recommended for performance
– wrappers now live in scikits.umfpack
– check-out the new scikits.suitesparse by Nathaniel Smith

11.2. Storage Schemes 380 11.3. Linear System Solvers 381

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Examples
"""
• import the whole module, and see its docstring: Solve a linear system
=======================
>>> help(sp.sparse.linalg.spsolve)
Help on function spsolve in module scipy.sparse.linalg._dsolve.linsolve: Construct a 1000x1000 lil_array and add some values to it, convert it
... to CSR format and solve A x = b for x:and solve a linear system with a
direct solver.
• both superlu and umfpack can be used (if the latter is installed) as follows: """
– prepare a linear system:
import numpy as np
>>> import numpy as np import scipy as sp
>>> mtx = sp.sparse.spdiags([[1, 2, 3, 4, 5], [6, 5, 8, 9, 10]], [0, 1], 5,␣ import matplotlib.pyplot as plt
˓→5, "csc")

>>> mtx.toarray() rng = np.random.default_rng(27446968)

array([[ 1, 5, 0, 0, 0],
[ 0, 2, 8, 0, 0], mtx = sp.sparse.lil_array((1000, 1000), dtype=np.float64)
[ 0, 0, 3, 9, 0], mtx[0, :100] = rng.random(100)
[ 0, 0, 0, 4, 10], mtx[1, 100:200] = mtx[[0], :100]
[ 0, 0, 0, 0, 5]]) mtx.setdiag(rng.random(1000))
>>> rhs = np.array([1, 2, 3, 4, 5], dtype=np.float32)
plt.clf()
– solve as single precision real: plt.spy(mtx, marker=".", markersize=2)
plt.show()
>>> mtx1 = mtx.astype(np.float32)
>>> x = sp.sparse.linalg.spsolve(mtx1, rhs, use_umfpack=False) mtx = mtx.tocsr()
>>> print(x) rhs = rng.random(1000)
[106. -21. 5.5 -1.5 1. ]
>>> print("Error: %s " % (mtx1 * x - rhs)) x = sp.sparse.linalg.spsolve(mtx, rhs)
Error: [0. 0. 0. 0. 0.]
print(f"residual: { np.linalg.norm(mtx @ x - rhs)!r} ")
– solve as double precision real:
• examples/direct_solve.py
>>> mtx2 = mtx.astype(np.float64)
>>> x = sp.sparse.linalg.spsolve(mtx2, rhs, use_umfpack=True)
>>> print(x) 11.3.2 Iterative Solvers
[106. -21. 5.5 -1.5 1. ] • the isolve module contains the following solvers:
>>> print("Error: %s " % (mtx2 * x - rhs))
Error: [0. 0. 0. 0. 0.] – bicg (BIConjugate Gradient)
– bicgstab (BIConjugate Gradient STABilized)
– solve as single precision complex:
– cg (Conjugate Gradient) - symmetric positive definite matrices only
>>> mtx1 = mtx.astype(np.complex64)
– cgs (Conjugate Gradient Squared)
>>> x = sp.sparse.linalg.spsolve(mtx1, rhs, use_umfpack=False)
>>> print(x) – gmres (Generalized Minimal RESidual)
[106. +0.j -21. +0.j 5.5+0.j -1.5+0.j 1. +0.j]
– minres (MINimum RESidual)
>>> print("Error: %s " % (mtx1 * x - rhs))
Error: [0.+0.j 0.+0.j 0.+0.j 0.+0.j 0.+0.j] – qmr (Quasi-Minimal Residual)

– solve as double precision complex: Common Parameters

>>> mtx2 = mtx.astype(np.complex128) • mandatory:
>>> x = sp.sparse.linalg.spsolve(mtx2, rhs, use_umfpack=True)
A
>>> print(x)
[{sparse array/matrix, dense array/matrix, LinearOperator}] The N-by-N matrix of the linear
[106. +0.j -21. +0.j 5.5+0.j -1.5+0.j 1. +0.j]
system.
>>> print("Error: %s " % (mtx2 * x - rhs))
Error: [0.+0.j 0.+0.j 0.+0.j 0.+0.j 0.+0.j] b
[{array, matrix}] Right hand side of the linear system. Has shape (N,) or (N,1).
• optional:

11.3. Linear System Solvers 382 11.3. Linear System Solvers 383
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

x0
"""
[{array, matrix}] Starting guess for the solution.
Compute eigenvectors and eigenvalues using a preconditioned eigensolver
tol =======================================================================
[float] Relative tolerance to achieve before terminating.
In this example Smoothed Aggregation (SA) is used to precondition
maxiter
the LOBPCG eigensolver on a two-dimensional Poisson problem with
[integer] Maximum number of iterations. Iteration will stop after maxiter steps even if the
Dirichlet boundary conditions.
specified tolerance has not been achieved.
"""
M
[{sparse array/matrix, dense array/matrix, LinearOperator}] Preconditioner for A. The pre- import numpy as np
conditioner should approximate the inverse of A. Effective preconditioning dramatically im- import scipy as sp
proves the rate of convergence, which implies that fewer iterations are needed to reach a given import matplotlib.pyplot as plt
error tolerance.
from pyamg import smoothed_aggregation_solver
callback
from pyamg.gallery import poisson
[function] User-supplied function to call after each iteration. It is called as callback(xk), where
xk is the current solution vector.
N = 100
K = 9
LinearOperator Class A = poisson((N, N), format="csr")
• common interface for performing matrix vector products
# create the AMG hierarchy
• useful abstraction that enables using dense and sparse matrices within the solvers, as well as ml = smoothed_aggregation_solver(A)
matrix-free solutions
• has shape and matvec() (+ some optional parameters) # initial approximation to the K eigenvectors
X = np.random.random((A.shape[0], K))
• example:
# preconditioner based on ml
>>> import numpy as np
M = ml.aspreconditioner()
>>> import scipy as sp
>>> def mv(v):
# compute eigenvalues and eigenvectors with LOBPCG
... return np.array([2 * v[0], 3 * v[1]])
W, V = sp.sparse.linalg.lobpcg(A, X, M=M, tol=1e-8, largest=False)
...
>>> A = sp.sparse.linalg.LinearOperator((2, 2), matvec=mv)
>>> A
# plot the eigenvectors
<2x2 _CustomLinearOperator with dtype=int8>
plt.figure(figsize=(9, 9))
>>> A.matvec(np.ones(2))
array([2., 3.])
for i in range(K):
>>> A * np.ones(2)
plt.subplot(3, 3, i + 1)
array([2., 3.])
plt.title(f"Eigenvector { i} ")
plt.pcolor(V[:, i].reshape(N, N))
A Few Notes on Preconditioning plt.axis("equal")
plt.axis("off")
• problem specific plt.show()
• often hard to develop
– examples/pyamg_with_lobpcg.py
• if not sure, try ILU
• example by Nils Wagner:
– available in scipy.sparse.linalg as spilu()
– examples/lobpcg_sakurai.py
11.3.3 Eigenvalue Problem Solvers • output:
The eigen module $ python examples/lobpcg_sakurai.py
Results by LOBPCG for n=2500
• arpack * a collection of Fortran77 subroutines designed to solve large scale eigenvalue problems
• lobpcg (Locally Optimal Block Preconditioned Conjugate Gradient Method) * works very well in [ 0.06250083 0.06250028 0.06250007]
combination with PyAMG * example by Nathan Bell:
Exact eigenvalues
(continues on next page)

11.3. Linear System Solvers 384 11.3. Linear System Solvers 385
 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

[ 0.06250005 0.0625002 0.06250044]

Elapsed time 7.01

CHAPTER 12
Image manipulation and processing
using NumPy and SciPy

Authors: Emmanuelle Gouillart, Gaël Varoquaux

This section addresses basic image manipulation and processing using the core scientific modules NumPy
and SciPy. Some of the operations covered by this tutorial may be useful for other kinds of multidimen-
sional array processing than image processing. In particular, the submodule scipy.ndimage provides
functions operating on n-dimensional NumPy arrays.

11.4 Other Interesting Packages

ã See also
• PyAMG For more advanced image processing and image-specific routines, see the tutorial scikit-image: image
– algebraic multigrid solvers processing, dedicated to the skimage module.

– https://github.com/pyamg/pyamg
• Pysparse Image = 2-D numerical array
– own sparse matrix classes
(or 3-D: CT, MRI, 2D + time; 4-D, . . . )
– matrix and eigenvalue problem solvers
Here, image == NumPy array np.array
– https://pysparse.sourceforge.net/
Tools used in this tutorial:
• numpy: basic array manipulation
• scipy: scipy.ndimage submodule dedicated to image processing (n-dimensional images). See the
documentation:

>>> import scipy as sp

11.4. Other Interesting Packages 386 387

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Common tasks in image processing:

• Input/Output, displaying images
• Basic manipulations: cropping, flipping, rotating, . . .
• Image filtering: denoising, sharpening
• Image segmentation: labeling pixels corresponding to different objects
• Classification
• Feature extraction
• Registration
• ...

Chapters contents

• Opening and writing to image files

• Displaying images
• Basic manipulations
– Statistical information
– Geometrical transformations
• Image filtering
Creating a NumPy array from an image file:
– Blurring/smoothing
>>> import imageio.v3 as iio
– Sharpening >>> face = sp.datasets.face()
– Denoising >>> iio.imwrite('face.png', face) # First we need to create the PNG file

– Mathematical morphology >>> face = iio.imread('face.png')

• Feature extraction >>> type(face)
<class 'numpy.ndarray'>
– Edge detection >>> face.shape, face.dtype
– Segmentation ((768, 1024, 3), dtype('uint8'))

• Measuring objects properties: scipy.ndimage.measurements dtype is uint8 for 8-bit images (0-255)
• Full code examples Opening raw files (camera, 3-D images)
• Examples for the image processing chapter
>>> face.tofile('face.raw') # Create raw file
>>> face_from_raw = np.fromfile('face.raw', dtype=np.uint8)
>>> face_from_raw.shape
(2359296,)
12.1 Opening and writing to image files >>> face_from_raw.shape = (768, 1024, 3)

Writing an array to a file: Need to know the shape and dtype of the image (how to separate data bytes).
import scipy as sp For large data, use np.memmap for memory mapping:
import imageio.v3 as iio
>>> face_memmap = np.memmap('face.raw', dtype=np.uint8, shape=(768, 1024, 3))
f = sp.datasets.face()
iio.imwrite("face.png", f) # uses the Image module (PIL) (data are read from the file, and not loaded into memory)
Working on a list of image files
import matplotlib.pyplot as plt
>>> rng = np.random.default_rng(27446968)
plt.imshow(f) >>> for i in range(10):
plt.show() ... im = rng.integers(0, 256, 10000, dtype=np.uint8).reshape((100, 100))
(continues on next page)

12.1. Opening and writing to image files 388 12.1. Opening and writing to image files 389
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

... iio.imwrite(f'random_{ i: 02d} .png', im)
>>> from glob import glob
>>> filelist = glob('random*.png')
>>> filelist.sort()

12.2 Displaying images

Use matplotlib and imshow to display an image inside a matplotlib figure:

>>> f = sp.datasets.face(gray=True) # retrieve a grayscale image

>>> import matplotlib.pyplot as plt
>>> plt.imshow(f, cmap=plt.cm.gray)
<matplotlib.image.AxesImage object at 0x...>

Increase contrast by setting min and max values:

>>> plt.imshow(f, cmap=plt.cm.gray, vmin=30, vmax=200)

<matplotlib.image.AxesImage object at 0x...>
>>> # Remove axes and ticks
>>> plt.axis('off') 12.3 Basic manipulations
(np.float64(-0.5), np.float64(1023.5), np.float64(767.5), np.float64(-0.5))
Images are arrays: use the whole numpy machinery.
Draw contour lines:

>>> plt.contour(f, [50, 200])

<matplotlib.contour.QuadContourSet ...>

For smooth intensity variations, use interpolation='bilinear'. For fine inspection of intensity varia-
>>> face = sp.datasets.face(gray=True)
tions, use interpolation='nearest':
>>> face[0, 40]
>>> plt.imshow(f[320:340, 510:530], cmap=plt.cm.gray, interpolation='bilinear') np.uint8(127)
<matplotlib.image.AxesImage object at 0x...> >>> # Slicing
>>> plt.imshow(f[320:340, 510:530], cmap=plt.cm.gray, interpolation='nearest') >>> face[10:13, 20:23]
<matplotlib.image.AxesImage object at 0x...> array([[141, 153, 145],
[133, 134, 125],
[ 96, 92, 94]], dtype=uint8)
>>> face[100:120] = 255
ã See also
>>>
More interpolation methods are in Matplotlib’s examples. >>> lx, ly = face.shape
>>> X, Y = np.ogrid[0:lx, 0:ly]
>>> mask = (X - lx / 2) ** 2 + (Y - ly / 2) ** 2 > lx * ly / 4
(continues on next page)

12.2. Displaying images 390 12.3. Basic manipulations 391

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> # Masks
>>> face[mask] = 0
>>> # Fancy indexing
>>> face[range(400), range(400)] = 255

12.3.2 Geometrical transformations

>>> face = sp.datasets.face(gray=True)
>>> lx, ly = face.shape
>>> # Cropping
>>> crop_face = face[lx // 4: - lx // 4, ly // 4: - ly // 4]
>>> # up <-> down flip
>>> flip_ud_face = np.flipud(face)
>>> # rotation
>>> rotate_face = sp.ndimage.rotate(face, 45)
>>> rotate_face_noreshape = sp.ndimage.rotate(face, 45, reshape=False)

12.3.1 Statistical information

>>> face = sp.datasets.face(gray=True)
>>> face.mean()
np.float64(113.48026784261067)
>>> face.max(), face.min() 12.4 Image filtering
(np.uint8(250), np.uint8(0))
Local filters: replace the value of pixels by a function of the values of neighboring pixels.
np.histogram Neighbourhood: square (choose size), disk, or more complicated structuring element.

Exercise

• Open as an array the scikit-image logo (https://scikit-image.org/_static/img/logo.png), or

an image that you have on your computer.
• Crop a meaningful part of the image, for example the python circle in the logo.
• Display the image array using matplotlib. Change the interpolation method and zoom to see 12.4.1 Blurring/smoothing
the difference.
Gaussian filter from scipy.ndimage:
• Transform your image to greyscale
• Increase the contrast of the image by changing its minimum and maximum values. Optional: >>> face = sp.datasets.face(gray=True)
use scipy.stats.scoreatpercentile (read the docstring!) to saturate 5% of the darkest >>> blurred_face = sp.ndimage.gaussian_filter(face, sigma=3)
pixels and 5% of the lightest pixels. >>> very_blurred = sp.ndimage.gaussian_filter(face, sigma=5)

• Save the array to two different file formats (png, jpg, tiff) Uniform filter

>>> local_mean = sp.ndimage.uniform_filter(face, size=11)

12.3. Basic manipulations 392 12.4. Image filtering 393

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

>>> med_denoised = sp.ndimage.median_filter(noisy, 3)

Median filter: better result for straight boundaries (low curvature):

12.4.2 Sharpening >>> im = np.zeros((20, 20))
Sharpen a blurred image: >>> im[5:-5, 5:-5] = 1
>>> im = sp.ndimage.distance_transform_bf(im)
>>> face = sp.datasets.face(gray=True).astype(float) >>> rng = np.random.default_rng()
>>> blurred_f = sp.ndimage.gaussian_filter(face, 3) >>> im_noise = im + 0.2 * rng.standard_normal(im.shape)
>>> im_med = sp.ndimage.median_filter(im_noise, 3)
increase the weight of edges by adding an approximation of the Laplacian:

>>> filter_blurred_f = sp.ndimage.gaussian_filter(blurred_f, 1)

>>> alpha = 30
>>> sharpened = blurred_f + alpha * (blurred_f - filter_blurred_f)

Other rank filter: scipy.ndimage.maximum_filter, scipy.ndimage.percentile_filter

Other local non-linear filters: Wiener (scipy.signal.wiener), etc.
Non-local filters

Exercise: denoising
12.4.3 Denoising
• Create a binary image (of 0s and 1s) with several objects (circles, ellipses, squares, or random
Noisy face:
shapes).
>>> f = sp.datasets.face(gray=True) • Add some noise (e.g., 20% of noise)
>>> f = f[230:290, 220:320]
>>> rng = np.random.default_rng() • Try two different denoising methods for denoising the image: gaussian filtering and median
>>> noisy = f + 0.4 * f.std() * rng.random(f.shape) filtering.
• Compare the histograms of the two different denoised images. Which one is the closest to the
A Gaussian filter smoothes the noise out. . . and the edges as well: histogram of the original (noise-free) image?
>>> gauss_denoised = sp.ndimage.gaussian_filter(noisy, 2)

Most local linear isotropic filters blur the image (scipy.ndimage.uniform_filter) ã See also

A median filter preserves better the edges: More denoising filters are available in skimage.denoising, see the scikit-image: image processing
tutorial.

12.4. Image filtering 394 12.4. Image filtering 395

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

12.4.4 Mathematical morphology (continued from previous page)

See wikipedia for a definition of mathematical morphology. [0, 0, 0, 0, 0, 0, 0],

[0, 0, 0, 0, 0, 0, 0],
Probe an image with a simple shape (a structuring element), and modify this image according to how [0, 0, 0, 0, 0, 0, 0]])
the shape locally fits or misses the image.
Structuring element:

>>> el = sp.ndimage.generate_binary_structure(2, 1)
>>> el
array([[False, True, False],
[ True, True, True],
[False, True, False]])
>>> el.astype(int)
array([[0, 1, 0],
[1, 1, 1],
[0, 1, 0]])

Dilation: maximum filter:

>>> a = np.zeros((5, 5))

>>> a[2, 2] = 1
>>> a
array([[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 1., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.]])
>>> sp.ndimage.binary_dilation(a).astype(a.dtype)
array([[0., 0., 0., 0., 0.],
[0., 0., 1., 0., 0.],
Erosion = minimum filter. Replace the value of a pixel by the minimal value covered by the structuring [0., 1., 1., 1., 0.],
element.: [0., 0., 1., 0., 0.],
[0., 0., 0., 0., 0.]])
>>> a = np.zeros((7,7), dtype=int)
>>> a[1:6, 2:5] = 1 Also works for grey-valued images:
>>> a
array([[0, 0, 0, 0, 0, 0, 0], >>> rng = np.random.default_rng(27446968)
[0, 0, 1, 1, 1, 0, 0], >>> im = np.zeros((64, 64))
[0, 0, 1, 1, 1, 0, 0], >>> x, y = (63*rng.random((2, 8))).astype(int)
[0, 0, 1, 1, 1, 0, 0], >>> im[x, y] = np.arange(8)
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0], >>> bigger_points = sp.ndimage.grey_dilation(im, size=(5, 5), structure=np.ones((5,␣
[0, 0, 0, 0, 0, 0, 0]]) ˓→5)))

>>> sp.ndimage.binary_erosion(a).astype(a.dtype)
array([[0, 0, 0, 0, 0, 0, 0], >>> square = np.zeros((16, 16))
[0, 0, 0, 0, 0, 0, 0], >>> square[4:-4, 4:-4] = 1
[0, 0, 0, 1, 0, 0, 0], >>> dist = sp.ndimage.distance_transform_bf(square)
[0, 0, 0, 1, 0, 0, 0], >>> dilate_dist = sp.ndimage.grey_dilation(dist, size=(3, 3), \
[0, 0, 0, 1, 0, 0, 0], ... structure=np.ones((3, 3)))
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
>>> # Erosion removes objects smaller than the structure
>>> sp.ndimage.binary_erosion(a, structure=np.ones((5,5))).astype(a.dtype)
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
(continues on next page)

12.4. Image filtering 396 12.4. Image filtering 397

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Opening: erosion + dilation:

>>> im = np.zeros((256, 256))
>>> a = np.zeros((5,5), dtype=int) >>> im[64:-64, 64:-64] = 1
>>> a[1:4, 1:4] = 1; a[4, 4] = 1 >>>
>>> a >>> im = sp.ndimage.rotate(im, 15, mode='constant')
array([[0, 0, 0, 0, 0], >>> im = sp.ndimage.gaussian_filter(im, 8)
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0], Use a gradient operator (Sobel) to find high intensity variations:
[0, 1, 1, 1, 0],
>>> sx = sp.ndimage.sobel(im, axis=0, mode='constant')
[0, 0, 0, 0, 1]])
>>> sy = sp.ndimage.sobel(im, axis=1, mode='constant')
>>> # Opening removes small objects
>>> sob = np.hypot(sx, sy)
>>> sp.ndimage.binary_opening(a, structure=np.ones((3,3))).astype(int)
array([[0, 0, 0, 0, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 0, 0, 0, 0]])
>>> # Opening can also smooth corners
>>> sp.ndimage.binary_opening(a).astype(int)
array([[0, 0, 0, 0, 0],
[0, 0, 1, 0, 0],
[0, 1, 1, 1, 0],
[0, 0, 1, 0, 0],
[0, 0, 0, 0, 0]])

Application: remove noise:

>>> square = np.zeros((32, 32)) 12.5.2 Segmentation

>>> square[10:-10, 10:-10] = 1
• Histogram-based segmentation (no spatial information)
>>> rng = np.random.default_rng(27446968)
>>> x, y = (32*rng.random((2, 20))).astype(int) >>> n = 10
>>> square[x, y] = 1 >>> l = 256
>>> im = np.zeros((l, l))
>>> open_square = sp.ndimage.binary_opening(square) >>> rng = np.random.default_rng(27446968)
>>> points = l*rng.random((2, n**2))
>>> eroded_square = sp.ndimage.binary_erosion(square) >>> im[(points[0]).astype(int), (points[1]).astype(int)] = 1
>>> reconstruction = sp.ndimage.binary_propagation(eroded_square, mask=square) >>> im = sp.ndimage.gaussian_filter(im, sigma=l/(4.*n))

>>> mask = (im > im.mean()).astype(float)

>>> mask += 0.1 * im
>>> img = mask + 0.2*rng.standard_normal(mask.shape)

>>> hist, bin_edges = np.histogram(img, bins=60)

>>> bin_centers = 0.5*(bin_edges[:-1] + bin_edges[1:])

>>> binary_img = img > 0.5

Closing: dilation + erosion
Use mathematical morphology to clean up the result:
Many other mathematical morphology operations: hit and miss transform, tophat, etc.
>>> # Remove small white regions
>>> open_img = sp.ndimage.binary_opening(binary_img)
12.5 Feature extraction >>> # Remove small black hole
>>> close_img = sp.ndimage.binary_closing(open_img)
12.5.1 Edge detection
Synthetic data:
Exercise

Check that reconstruction operations (erosion + propagation) produce a better result than open-

12.5. Feature extraction 398 12.5. Feature extraction 399

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

>>> from sklearn.cluster import spectral_clustering

>>> l = 100
>>> x, y = np.indices((l, l))

>>> center1 = (28, 24)

>>> center2 = (40, 50)
>>> center3 = (67, 58)
>>> center4 = (24, 70)
>>> radius1, radius2, radius3, radius4 = 16, 14, 15, 14

>>> circle1 = (x - center1[0])**2 + (y - center1[1])**2 < radius1**2

>>> circle2 = (x - center2[0])**2 + (y - center2[1])**2 < radius2**2
>>> circle3 = (x - center3[0])**2 + (y - center3[1])**2 < radius3**2
>>> circle4 = (x - center4[0])**2 + (y - center4[1])**2 < radius4**2

>>> # 4 circles
>>> img = circle1 + circle2 + circle3 + circle4
>>> mask = img.astype(bool)
>>> img = img.astype(float)

>>> rng = np.random.default_rng()

>>> img += 1 + 0.2*rng.standard_normal(img.shape)
>>> # Convert the image into a graph with the value of the gradient on
>>> # the edges.
>>> graph = image.img_to_graph(img, mask=mask)
ing/closing:
>>> eroded_img = sp.ndimage.binary_erosion(binary_img)
>>> # Take a decreasing function of the gradient: we take it weakly
>>> reconstruct_img = sp.ndimage.binary_propagation(eroded_img, mask=binary_img)
>>> # dependent from the gradient the segmentation is close to a voronoi
>>> tmp = np.logical_not(reconstruct_img)
>>> graph.data = np.exp(-graph.data/graph.data.std())
>>> eroded_tmp = sp.ndimage.binary_erosion(tmp)
>>> reconstruct_final = np.logical_not(sp.ndimage.binary_propagation(eroded_tmp,␣
>>> labels = spectral_clustering(graph, n_clusters=4, eigen_solver='arpack')
˓→mask=tmp))
>>> label_im = -np.ones(mask.shape)
>>> np.abs(mask - close_img).mean()
>>> label_im[mask] = labels
np.float64(0.00640699...)
>>> np.abs(mask - reconstruct_final).mean()
np.float64(0.00082232...)

Exercise

Check how a first denoising step (e.g. with a median filter) modifies the histogram, and check that
the resulting histogram-based segmentation is more accurate.

ã See also

More advanced segmentation algorithms are found in the scikit-image: see scikit-image: image
processing.

ã See also
12.6 Measuring objects properties: scipy.ndimage.measurements
Other Scientific Packages provide algorithms that can be useful for image processing. In this example,
we use the spectral clustering function of the scikit-learn in order to segment glued objects. Synthetic data:
>>> from sklearn.feature_extraction import image

12.5. Feature extraction 400 12.6. Measuring objects properties: scipy.ndimage.measurements 401
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

>>> n = 10
>>> l = 256
>>> im = np.zeros((l, l))
>>> rng = np.random.default_rng(27446968)
>>> points = l * rng.random((2, n**2))
>>> im[(points[0]).astype(int), (points[1]).astype(int)] = 1
>>> im = sp.ndimage.gaussian_filter(im, sigma=l/(4.*n))
>>> mask = im > im.mean()

• Analysis of connected components

Label connected components: scipy.dimage.label:

>>> label_im, nb_labels = sp.ndimage.label(mask)

>>> nb_labels # how many regions?
28
>>> plt.imshow(label_im)
<matplotlib.image.AxesImage object at 0x...>

Compute size, mean_value, etc. of each region:

>>> sizes = sp.ndimage.sum(mask, label_im, range(nb_labels + 1))

>>> mean_vals = sp.ndimage.sum(im, label_im, range(1, nb_labels + 1))

Clean up small connect components:

>>> mask_size = sizes < 1000

>>> remove_pixel = mask_size[label_im]
>>> remove_pixel.shape
(256, 256)
>>> label_im[remove_pixel] = 0
>>> plt.imshow(label_im)
<matplotlib.image.AxesImage object at 0x...>

Now reassign labels with np.searchsorted:

>>> labels = np.unique(label_im)

>>> label_im = np.searchsorted(labels, label_im)

Find region of interest enclosing object:

>>> slice_x, slice_y = sp.ndimage.find_objects(label_im)[3]

>>> roi = im[slice_x, slice_y]
>>> plt.imshow(roi)
<matplotlib.image.AxesImage object at 0x...>

12.6. Measuring objects properties: scipy.ndimage.measurements 402 12.6. Measuring objects properties: scipy.ndimage.measurements 403
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Other spatial measures: scipy.ndimage.center_of_mass, scipy.ndimage.maximum_position, etc.

Can be used outside the limited scope of segmentation applications.
Example: block mean:

>>> f = sp.datasets.face(gray=True)
>>> sx, sy = f.shape
>>> X, Y = np.ogrid[0:sx, 0:sy]
>>> regions = (sy//6) * (X//4) + (Y//6) # note that we use broadcasting
>>> block_mean = sp.ndimage.mean(f, labels=regions, index=np.arange(1,
... regions.max() +1))
>>> block_mean.shape = (sx // 4, sy // 6)

(continued from previous page)

>>>
>>> def granulometry(data, sizes=None):
... s = max(data.shape)
... if sizes is None:
... sizes = range(1, s/2, 2)
... granulo = [sp.ndimage.binary_opening(data, \
... structure=disk_structure(n)).sum() for n in sizes]
... return granulo
...
>>>
>>> rng = np.random.default_rng(27446968)
When regions are regular blocks, it is more efficient to use stride tricks (Example: fake dimensions with >>> n = 10
strides). >>> l = 256
>>> im = np.zeros((l, l))
Non-regularly-spaced blocks: radial mean: >>> points = l*rng.random((2, n**2))
>>> im[(points[0]).astype(int), (points[1]).astype(int)] = 1
>>> sx, sy = f.shape
>>> im = sp.ndimage.gaussian_filter(im, sigma=l/(4.*n))
>>> X, Y = np.ogrid[0:sx, 0:sy]
>>>
>>> r = np.hypot(X - sx/2, Y - sy/2)
>>> mask = im > im.mean()
>>> rbin = (20* r/r.max()).astype(int)
>>>
>>> radial_mean = sp.ndimage.mean(f, labels=rbin, index=np.arange(1, rbin.max() +1))
>>> granulo = granulometry(mask, sizes=np.arange(2, 19, 4))
• Other measures
Correlation function, Fourier/wavelet spectrum, etc. 12.7 Full code examples
One example with mathematical morphology: granulometry
12.8 Examples for the image processing chapter
>>> def disk_structure(n):
... struct = np.zeros((2 * n + 1, 2 * n + 1))
12.8.1 Displaying a Raccoon Face
... x, y = np.indices((2 * n + 1, 2 * n + 1))
... mask = (x - n)**2 + (y - n)**2 <= n**2 Small example to plot a raccoon face.
... struct[mask] = 1
... return struct.astype(bool)
...
(continues on next page)

12.6. Measuring objects properties: scipy.ndimage.measurements 404 12.7. Full code examples 405
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

12.8.2 Image interpolation

The example demonstrates image interpolation on a Raccoon face.

import scipy as sp
import matplotlib.pyplot as plt

f = sp.datasets.face(gray=True)

plt.figure(figsize=(8, 4))

plt.subplot(1, 2, 1)
plt.imshow(f[320:340, 510:530], cmap="gray")
plt.axis("off")

plt.subplot(1, 2, 2)
plt.imshow(f[320:340, 510:530], cmap="gray", interpolation="nearest")
plt.axis("off")

plt.subplots_adjust(wspace=0.02, hspace=0.02, top=1, bottom=0, left=0, right=1)

plt.show()

Total running time of the script: (0 minutes 0.173 seconds)

import scipy as sp
import imageio.v3 as iio 12.8.3 Plot the block mean of an image
An example showing how to use broad-casting to plot the mean of blocks of an image.
f = sp.datasets.face()
iio.imwrite("face.png", f) # uses the Image module (PIL)

import matplotlib.pyplot as plt

plt.imshow(f)
plt.show()

Total running time of the script: (0 minutes 0.539 seconds)

12.8. Examples for the image processing chapter 406 12.8. Examples for the image processing chapter 407
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

import numpy as np
import scipy as sp
import matplotlib.pyplot as plt

face = sp.datasets.face(gray=True)
face[10:13, 20:23]
face[100:120] = 255

lx, ly = face.shape
X, Y = np.ogrid[0:lx, 0:ly]
mask = (X - lx / 2) ** 2 + (Y - ly / 2) ** 2 > lx * ly / 4
face[mask] = 0
import numpy as np face[range(400), range(400)] = 255
import scipy as sp
import matplotlib.pyplot as plt plt.figure(figsize=(3, 3))
plt.axes((0, 0, 1, 1))
f = sp.datasets.face(gray=True) plt.imshow(face, cmap="gray")
sx, sy = f.shape plt.axis("off")
X, Y = np.ogrid[0:sx, 0:sy]
plt.show()
regions = sy // 6 * (X // 4) + Y // 6
block_mean = sp.ndimage.mean(f, labels=regions, index=np.arange(1, regions.max() + 1)) Total running time of the script: (0 minutes 0.200 seconds)
block_mean.shape = (sx // 4, sy // 6)

plt.figure(figsize=(5, 5))
plt.imshow(block_mean, cmap="gray") 12.8.5 Radial mean
plt.axis("off") This example shows how to do a radial mean with scikit-image.
plt.show()

Total running time of the script: (0 minutes 0.196 seconds)

12.8.4 Image manipulation and NumPy arrays

This example shows how to do image manipulation using common NumPy arrays tricks.

12.8. Examples for the image processing chapter 408 12.8. Examples for the image processing chapter 409
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

12.8.6 Display a Raccoon Face

An example that displays a raccoon face with matplotlib.

import scipy as sp
import matplotlib.pyplot as plt

f = sp.datasets.face(gray=True)

plt.figure(figsize=(10, 3.6))

plt.subplot(131)
plt.imshow(f, cmap="gray")

plt.subplot(132)
plt.imshow(f, cmap="gray", vmin=30, vmax=200)
plt.axis("off")
import numpy as np
import scipy as sp plt.subplot(133)
import matplotlib.pyplot as plt plt.imshow(f, cmap="gray")
plt.contour(f, [50, 200])
f = sp.datasets.face(gray=True) plt.axis("off")
sx, sy = f.shape
X, Y = np.ogrid[0:sx, 0:sy] plt.subplots_adjust(wspace=0, hspace=0.0, top=0.99, bottom=0.01, left=0.05, right=0.
˓→99)

plt.show()
r = np.hypot(X - sx / 2, Y - sy / 2)
Total running time of the script: (0 minutes 0.411 seconds)
rbin = (20 * r / r.max()).astype(int)
radial_mean = sp.ndimage.mean(f, labels=rbin, index=np.arange(1, rbin.max() + 1))

plt.figure(figsize=(5, 5))
12.8.7 Image sharpening
plt.axes((0, 0, 1, 1)) This example shows how to sharpen an image in noiseless situation by applying the filter inverse to the
plt.imshow(rbin, cmap="nipy_spectral") blur.
plt.axis("off")

plt.show()

Total running time of the script: (0 minutes 0.226 seconds)

12.8. Examples for the image processing chapter 410 12.8. Examples for the image processing chapter 411
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

import scipy as sp import scipy as sp

import matplotlib.pyplot as plt import matplotlib.pyplot as plt

f = sp.datasets.face(gray=True).astype(float) face = sp.datasets.face(gray=True)

blurred_f = sp.ndimage.gaussian_filter(f, 3) blurred_face = sp.ndimage.gaussian_filter(face, sigma=3)
very_blurred = sp.ndimage.gaussian_filter(face, sigma=5)
filter_blurred_f = sp.ndimage.gaussian_filter(blurred_f, 1) local_mean = sp.ndimage.uniform_filter(face, size=11)

alpha = 30 plt.figure(figsize=(9, 3))

sharpened = blurred_f + alpha * (blurred_f - filter_blurred_f) plt.subplot(131)
plt.imshow(blurred_face, cmap="gray")
plt.figure(figsize=(12, 4)) plt.axis("off")
plt.subplot(132)
plt.subplot(131) plt.imshow(very_blurred, cmap="gray")
plt.imshow(f, cmap="gray") plt.axis("off")
plt.axis("off") plt.subplot(133)
plt.subplot(132) plt.imshow(local_mean, cmap="gray")
plt.imshow(blurred_f, cmap="gray") plt.axis("off")
plt.axis("off")
plt.subplot(133) plt.subplots_adjust(wspace=0, hspace=0.0, top=0.99, bottom=0.01, left=0.01, right=0.
plt.imshow(sharpened, cmap="gray") ˓→99)

plt.axis("off")
plt.show()
plt.tight_layout()
plt.show() Total running time of the script: (0 minutes 0.361 seconds)

Total running time of the script: (0 minutes 0.531 seconds)

12.8.9 Synthetic data

The example generates and displays simple synthetic data.
12.8.8 Blurring of images
An example showing various processes that blur an image.

12.8. Examples for the image processing chapter 412 12.8. Examples for the image processing chapter 413
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

import numpy as np
import scipy as sp rng = np.random.default_rng(27446968)
import matplotlib.pyplot as plt x, y = (32 * rng.random((2, 20))).astype(int)
square[x, y] = 1
rng = np.random.default_rng(27446968)
n = 10 open_square = sp.ndimage.binary_opening(square)
l = 256
im = np.zeros((l, l)) eroded_square = sp.ndimage.binary_erosion(square)
points = l * rng.random((2, n**2)) reconstruction = sp.ndimage.binary_propagation(eroded_square, mask=square)
im[(points[0]).astype(int), (points[1]).astype(int)] = 1
im = sp.ndimage.gaussian_filter(im, sigma=l / (4.0 * n)) plt.figure(figsize=(9.5, 3))
plt.subplot(131)
mask = im > im.mean() plt.imshow(square, cmap="gray", interpolation="nearest")
plt.axis("off")
label_im, nb_labels = sp.ndimage.label(mask) plt.subplot(132)
plt.imshow(open_square, cmap="gray", interpolation="nearest")
plt.figure(figsize=(9, 3)) plt.axis("off")
plt.subplot(133)
plt.subplot(131) plt.imshow(reconstruction, cmap="gray", interpolation="nearest")
plt.imshow(im) plt.axis("off")
plt.axis("off")
plt.subplot(132) plt.subplots_adjust(wspace=0, hspace=0.02, top=0.99, bottom=0.01, left=0.01, right=0.
˓→99)
plt.imshow(mask, cmap="gray")
plt.axis("off") plt.show()
plt.subplot(133)
plt.imshow(label_im, cmap="nipy_spectral") Total running time of the script: (0 minutes 0.036 seconds)
plt.axis("off")

plt.subplots_adjust(wspace=0.02, hspace=0.02, top=1, bottom=0, left=0, right=1)

plt.show() 12.8.11 Image denoising
This example demoes image denoising on a Raccoon face.
Total running time of the script: (0 minutes 0.080 seconds)

12.8.10 Opening, erosion, and propagation

This example shows simple operations of mathematical morphology.

import numpy as np
import scipy as sp
import matplotlib.pyplot as plt

rng = np.random.default_rng(27446968)

f = sp.datasets.face(gray=True)
f = f[230:290, 220:320]

import numpy as np noisy = f + 0.4 * f.std() * rng.random(f.shape)

import scipy as sp
import matplotlib.pyplot as plt gauss_denoised = sp.ndimage.gaussian_filter(noisy, 2)
med_denoised = sp.ndimage.median_filter(noisy, 3)
square = np.zeros((32, 32))
square[10:-10, 10:-10] = 1
(continues on next page)
(continues on next page)

12.8. Examples for the image processing chapter 414 12.8. Examples for the image processing chapter 415
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

plt.figure(figsize=(12, 2.8)) plt.subplot(153)
plt.imshow(flip_ud_face, cmap="gray")
plt.subplot(131) plt.axis("off")
plt.imshow(noisy, cmap="gray", vmin=40, vmax=220) plt.subplot(154)
plt.axis("off") plt.imshow(rotate_face, cmap="gray")
plt.title("noisy", fontsize=20) plt.axis("off")
plt.subplot(132) plt.subplot(155)
plt.imshow(gauss_denoised, cmap="gray", vmin=40, vmax=220) plt.imshow(rotate_face_noreshape, cmap="gray")
plt.axis("off") plt.axis("off")
plt.title("Gaussian filter", fontsize=20)
plt.subplot(133) plt.subplots_adjust(wspace=0.02, hspace=0.3, top=1, bottom=0.1, left=0, right=1)
plt.imshow(med_denoised, cmap="gray", vmin=40, vmax=220)
plt.axis("off") plt.show()
plt.title("Median filter", fontsize=20)
Total running time of the script: (0 minutes 0.587 seconds)
plt.subplots_adjust(wspace=0.02, hspace=0.02, top=0.9, bottom=0, left=0, right=1)
plt.show()

Total running time of the script: (0 minutes 0.215 seconds) 12.8.13 Total Variation denoising
This example demoes Total-Variation (TV) denoising on a Raccoon face.

12.8.12 Geometrical transformations

This examples demos some simple geometrical transformations on a Raccoon face.

import numpy as np
import scipy as sp
import matplotlib.pyplot as plt
import numpy as np
import scipy as sp
from skimage.restoration import denoise_tv_chambolle
import matplotlib.pyplot as plt
rng = np.random.default_rng(27446968)
face = sp.datasets.face(gray=True)
lx, ly = face.shape
f = sp.datasets.face(gray=True)
# Cropping
f = f[230:290, 220:320]
crop_face = face[lx // 4 : -lx // 4, ly // 4 : -ly // 4]
# up <-> down flip
noisy = f + 0.4 * f.std() * rng.random(f.shape)
flip_ud_face = np.flipud(face)
# rotation
tv_denoised = denoise_tv_chambolle(noisy, weight=10)
rotate_face = sp.ndimage.rotate(face, 45)
rotate_face_noreshape = sp.ndimage.rotate(face, 45, reshape=False)
plt.figure(figsize=(12, 2.8))
plt.figure(figsize=(12.5, 2.5))
plt.subplot(131)
plt.imshow(noisy, cmap="gray", vmin=40, vmax=220)
plt.subplot(151)
plt.axis("off")
plt.imshow(face, cmap="gray")
plt.title("noisy", fontsize=20)
plt.axis("off")
plt.subplot(132)
plt.subplot(152)
plt.imshow(tv_denoised, cmap="gray", vmin=40, vmax=220)
plt.imshow(crop_face, cmap="gray")
plt.axis("off")
plt.axis("off")
(continues on next page)
(continues on next page)

12.8. Examples for the image processing chapter 416 12.8. Examples for the image processing chapter 417
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

plt.title("TV denoising", fontsize=20) label_im[remove_pixel] = 0
labels = np.unique(label_im)
tv_denoised = denoise_tv_chambolle(noisy, weight=50) label_clean = np.searchsorted(labels, label_im)
plt.subplot(133)
plt.imshow(tv_denoised, cmap="gray", vmin=40, vmax=220)
plt.axis("off") plt.figure(figsize=(6, 3))
plt.title("(more) TV denoising", fontsize=20)
plt.subplot(121)
plt.subplots_adjust(wspace=0.02, hspace=0.02, top=0.9, bottom=0, left=0, right=1) plt.imshow(label_im, cmap="nipy_spectral")
plt.show() plt.axis("off")
plt.subplot(122)
Total running time of the script: (0 minutes 0.229 seconds) plt.imshow(label_clean, vmax=nb_labels, cmap="nipy_spectral")
plt.axis("off")

plt.subplots_adjust(wspace=0.01, hspace=0.01, top=1, bottom=0, left=0, right=1)

12.8.14 Measurements from images plt.show()
This examples shows how to measure quantities from various images.
Total running time of the script: (0 minutes 0.042 seconds)

12.8.15 Find the bounding box of an object

This example shows how to extract the bounding box of the largest object

import numpy as np import numpy as np

import scipy as sp import scipy as sp
import matplotlib.pyplot as plt import matplotlib.pyplot as plt

rng = np.random.default_rng(27446968) rng = np.random.default_rng(27446968)

n = 10 n = 10
l = 256 l = 256
im = np.zeros((l, l)) im = np.zeros((l, l))
points = l * rng.random((2, n**2)) points = l * rng.random((2, n**2))
im[(points[0]).astype(int), (points[1]).astype(int)] = 1 im[(points[0]).astype(int), (points[1]).astype(int)] = 1
im = sp.ndimage.gaussian_filter(im, sigma=l / (4.0 * n)) im = sp.ndimage.gaussian_filter(im, sigma=l / (4.0 * n))

mask = im > im.mean() mask = im > im.mean()

label_im, nb_labels = sp.ndimage.label(mask) label_im, nb_labels = sp.ndimage.label(mask)

sizes = sp.ndimage.sum(mask, label_im, range(nb_labels + 1)) # Find the largest connected component
mask_size = sizes < 1000 sizes = sp.ndimage.sum(mask, label_im, range(nb_labels + 1))
remove_pixel = mask_size[label_im] mask_size = sizes < 1000
(continues on next page) (continues on next page)

12.8. Examples for the image processing chapter 418 12.8. Examples for the image processing chapter 419
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

remove_pixel = mask_size[label_im] plt.subplot(142)
label_im[remove_pixel] = 0 plt.imshow(im_noise, interpolation="nearest", vmin=0, vmax=5)
labels = np.unique(label_im) plt.axis("off")
label_im = np.searchsorted(labels, label_im) plt.title("Noisy image", fontsize=20)
plt.subplot(143)
# Now that we have only one connected component, extract it's bounding box plt.imshow(im_med, interpolation="nearest", vmin=0, vmax=5)
slice_x, slice_y = sp.ndimage.find_objects(label_im == 4)[0] plt.axis("off")
roi = im[slice_x, slice_y] plt.title("Median filter", fontsize=20)
plt.subplot(144)
plt.figure(figsize=(4, 2)) plt.imshow(np.abs(im - im_med), cmap="hot", interpolation="nearest")
plt.axes((0, 0, 1, 1)) plt.axis("off")
plt.imshow(roi) plt.title("Error", fontsize=20)
plt.axis("off")

plt.show() plt.subplots_adjust(wspace=0.02, hspace=0.02, top=0.9, bottom=0, left=0, right=1)

Total running time of the script: (0 minutes 0.017 seconds) plt.show()

Total running time of the script: (0 minutes 0.119 seconds)

12.8.16 Denoising an image with the median filter

This example shows the original image, the noisy image, the denoised one (with the median filter) and
12.8.17 Histogram segmentation
the difference between the two.
This example does simple histogram analysis to perform segmentation.

import numpy as np
import scipy as sp
import matplotlib.pyplot as plt
import numpy as np
rng = np.random.default_rng(27446968) import scipy as sp
import matplotlib.pyplot as plt
im = np.zeros((20, 20))
im[5:-5, 5:-5] = 1 rng = np.random.default_rng(27446968)
im = sp.ndimage.distance_transform_bf(im) n = 10
im_noise = im + 0.2 * rng.normal(size=im.shape) l = 256
im = np.zeros((l, l))
im_med = sp.ndimage.median_filter(im_noise, 3) points = l * rng.random((2, n**2))
im[(points[0]).astype(int), (points[1]).astype(int)] = 1
plt.figure(figsize=(16, 5)) im = sp.ndimage.gaussian_filter(im, sigma=l / (4.0 * n))

plt.subplot(141) mask = (im > im.mean()).astype(float)

plt.imshow(im, interpolation="nearest")
plt.axis("off") mask += 0.1 * im
plt.title("Original image", fontsize=20)
(continues on next page) (continues on next page)

12.8. Examples for the image processing chapter 420 12.8. Examples for the image processing chapter 421
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

img = mask + 0.2 * rng.normal(size=mask.shape)
plt.figure(figsize=(12.5, 3))
hist, bin_edges = np.histogram(img, bins=60) plt.subplot(141)
bin_centers = 0.5 * (bin_edges[:-1] + bin_edges[1:]) plt.imshow(im, interpolation="nearest", cmap="nipy_spectral")
plt.axis("off")
binary_img = img > 0.5 plt.subplot(142)
plt.imshow(bigger_points, interpolation="nearest", cmap="nipy_spectral")
plt.figure(figsize=(11, 4)) plt.axis("off")
plt.subplot(143)
plt.subplot(131) plt.imshow(dist, interpolation="nearest", cmap="nipy_spectral")
plt.imshow(img) plt.axis("off")
plt.axis("off") plt.subplot(144)
plt.subplot(132) plt.imshow(dilate_dist, interpolation="nearest", cmap="nipy_spectral")
plt.plot(bin_centers, hist, lw=2) plt.axis("off")
plt.axvline(0.5, color="r", ls="--", lw=2)
plt.text(0.57, 0.8, "histogram", fontsize=20, transform=plt.gca().transAxes) plt.subplots_adjust(wspace=0, hspace=0.02, top=0.99, bottom=0.01, left=0.01, right=0.
plt.yticks([]) ˓→99)

plt.subplot(133) plt.show()
plt.imshow(binary_img, cmap="gray", interpolation="nearest")
plt.axis("off") Total running time of the script: (0 minutes 0.049 seconds)

plt.subplots_adjust(wspace=0.02, hspace=0.3, top=1, bottom=0.1, left=0, right=1)

plt.show()
12.8.19 Finding edges with Sobel filters
Total running time of the script: (0 minutes 0.103 seconds) The Sobel filter is one of the simplest way of finding edges.

12.8.18 Greyscale dilation

This example illustrates greyscale mathematical morphology.

import numpy as np
import scipy as sp
import numpy as np import matplotlib.pyplot as plt
import scipy as sp
import matplotlib.pyplot as plt rng = np.random.default_rng(27446968)

im = np.zeros((64, 64)) im = np.zeros((256, 256))

rng = np.random.default_rng(27446968) im[64:-64, 64:-64] = 1
x, y = (63 * rng.random((2, 8))).astype(int)
im[x, y] = np.arange(8) im = sp.ndimage.rotate(im, 15, mode="constant")
im = sp.ndimage.gaussian_filter(im, 8)
bigger_points = sp.ndimage.grey_dilation(im, size=(5, 5), structure=np.ones((5, 5)))
sx = sp.ndimage.sobel(im, axis=0, mode="constant")
square = np.zeros((16, 16)) sy = sp.ndimage.sobel(im, axis=1, mode="constant")
square[4:-4, 4:-4] = 1 sob = np.hypot(sx, sy)
dist = sp.ndimage.distance_transform_bf(square)
dilate_dist = sp.ndimage.grey_dilation(dist, size=(3, 3), structure=np.ones((3, 3))) plt.figure(figsize=(16, 5))
(continues on next page) (continues on next page)

12.8. Examples for the image processing chapter 422 12.8. Examples for the image processing chapter 423
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

plt.subplot(141) im[(points[0]).astype(int), (points[1]).astype(int)] = 1
plt.imshow(im, cmap="gray") im = sp.ndimage.gaussian_filter(im, sigma=l / (4.0 * n))
plt.axis("off")
plt.title("square", fontsize=20) mask = (im > im.mean()).astype(float)
plt.subplot(142)
plt.imshow(sx)
plt.axis("off") img = mask + 0.3 * rng.normal(size=mask.shape)
plt.title("Sobel (x direction)", fontsize=20)
plt.subplot(143) binary_img = img > 0.5
plt.imshow(sob)
plt.axis("off") # Remove small white regions
plt.title("Sobel filter", fontsize=20) open_img = sp.ndimage.binary_opening(binary_img)
# Remove small black hole
im += 0.07 * rng.random(im.shape) close_img = sp.ndimage.binary_closing(open_img)

sx = sp.ndimage.sobel(im, axis=0, mode="constant") plt.figure(figsize=(12, 3))

sy = sp.ndimage.sobel(im, axis=1, mode="constant")
sob = np.hypot(sx, sy) l = 128

plt.subplot(144) plt.subplot(141)
plt.imshow(sob) plt.imshow(binary_img[:l, :l], cmap="gray")
plt.axis("off") plt.axis("off")
plt.title("Sobel for noisy image", fontsize=20) plt.subplot(142)
plt.imshow(open_img[:l, :l], cmap="gray")
plt.axis("off")
plt.subplots_adjust(wspace=0.02, hspace=0.02, top=1, bottom=0, left=0, right=0.9) plt.subplot(143)
plt.imshow(close_img[:l, :l], cmap="gray")
plt.show() plt.axis("off")
plt.subplot(144)
Total running time of the script: (0 minutes 0.196 seconds) plt.imshow(mask[:l, :l], cmap="gray")
plt.contour(close_img[:l, :l], [0.5], linewidths=2, colors="r")
plt.axis("off")

12.8.20 Cleaning segmentation with mathematical morphology plt.subplots_adjust(wspace=0.02, hspace=0.3, top=1, bottom=0.1, left=0, right=1)
An example showing how to clean segmentation with mathematical morphology: removing small regions
and holes. plt.show()

Total running time of the script: (0 minutes 0.076 seconds)

12.8.21 Segmentation with Gaussian mixture models

This example performs a Gaussian mixture model analysis of the image histogram to find the right
thresholds for separating foreground from background.

import numpy as np
import scipy as sp
import matplotlib.pyplot as plt

rng = np.random.default_rng(27446968)
n = 10
l = 256
im = np.zeros((l, l))
points = l * rng.random((2, n**2))
(continues on next page)

12.8. Examples for the image processing chapter 424 12.8. Examples for the image processing chapter 425
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

plt.show()

Total running time of the script: (0 minutes 0.707 seconds)

12.8.22 Watershed segmentation

This example shows how to do segmentation with watershed.

import numpy as np
import scipy as sp
import matplotlib.pyplot as plt
from sklearn.mixture import GaussianMixture

rng = np.random.default_rng(27446968)
n = 10
l = 256
im = np.zeros((l, l))
points = l * rng.random((2, n**2))
im[(points[0]).astype(int), (points[1]).astype(int)] = 1
im = sp.ndimage.gaussian_filter(im, sigma=l / (4.0 * n)) import numpy as np
from skimage.segmentation import watershed
mask = (im > im.mean()).astype(float) from skimage.feature import peak_local_max
import matplotlib.pyplot as plt
import scipy as sp
img = mask + 0.3 * rng.normal(size=mask.shape)
# Generate an initial image with two overlapping circles
hist, bin_edges = np.histogram(img, bins=60) x, y = np.indices((80, 80))
bin_centers = 0.5 * (bin_edges[:-1] + bin_edges[1:]) x1, y1, x2, y2 = 28, 28, 44, 52
r1, r2 = 16, 20
classif = GaussianMixture(n_components=2) mask_circle1 = (x - x1) ** 2 + (y - y1) ** 2 < r1**2
classif.fit(img.reshape((img.size, 1))) mask_circle2 = (x - x2) ** 2 + (y - y2) ** 2 < r2**2
image = np.logical_or(mask_circle1, mask_circle2)
threshold = np.mean(classif.means_) # Now we want to separate the two objects in image
binary_img = img > threshold # Generate the markers as local maxima of the distance
# to the background
distance = sp.ndimage.distance_transform_edt(image)
plt.figure(figsize=(11, 4)) peak_idx = peak_local_max(distance, footprint=np.ones((3, 3)), labels=image)
peak_mask = np.zeros_like(distance, dtype=bool)
plt.subplot(131) peak_mask[tuple(peak_idx.T)] = True
plt.imshow(img) markers = sp.ndimage.label(peak_mask)[0]
plt.axis("off") labels = watershed(-distance, markers, mask=image)
plt.subplot(132)
plt.plot(bin_centers, hist, lw=2) plt.figure(figsize=(9, 3.5))
plt.axvline(0.5, color="r", ls="--", lw=2) plt.subplot(131)
plt.text(0.57, 0.8, "histogram", fontsize=20, transform=plt.gca().transAxes) plt.imshow(image, cmap="gray", interpolation="nearest")
plt.yticks([]) plt.axis("off")
plt.subplot(133) plt.subplot(132)
plt.imshow(binary_img, cmap="gray", interpolation="nearest") plt.imshow(-distance, interpolation="nearest")
plt.axis("off") plt.axis("off")
plt.subplot(133)
plt.subplots_adjust(wspace=0.02, hspace=0.3, top=1, bottom=0.1, left=0, right=1) (continues on next page)
(continues on next page)

12.8. Examples for the image processing chapter 426 12.8. Examples for the image processing chapter 427
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

plt.imshow(labels, cmap="nipy_spectral", interpolation="nearest") im[(points[0]).astype(int), (points[1]).astype(int)] = 1
plt.axis("off") im = sp.ndimage.gaussian_filter(im, sigma=l / (4.0 * n))

plt.subplots_adjust(hspace=0.01, wspace=0.01, top=1, bottom=0, left=0, right=1) mask = im > im.mean()

plt.show()
granulo = granulometry(mask, sizes=np.arange(2, 19, 4))
Total running time of the script: (0 minutes 0.059 seconds)
plt.figure(figsize=(6, 2.2))

plt.subplot(121)
12.8.23 Granulometry plt.imshow(mask, cmap="gray")
opened = sp.ndimage.binary_opening(mask, structure=disk_structure(10))
This example performs a simple granulometry analysis.
opened_more = sp.ndimage.binary_opening(mask, structure=disk_structure(14))
plt.contour(opened, [0.5], colors="b", linewidths=2)
plt.contour(opened_more, [0.5], colors="r", linewidths=2)
plt.axis("off")
plt.subplot(122)
plt.plot(np.arange(2, 19, 4), granulo, "ok", ms=8)

plt.subplots_adjust(wspace=0.02, hspace=0.15, top=0.95, bottom=0.15, left=0, right=0.

˓→95)

plt.show()

Total running time of the script: (0 minutes 0.244 seconds)

import numpy as np
import scipy as sp 12.8.24 Segmentation with spectral clustering
import matplotlib.pyplot as plt This example uses spectral clustering to do segmentation.

import numpy as np
def disk_structure(n): import matplotlib.pyplot as plt
struct = np.zeros((2 * n + 1, 2 * n + 1))
x, y = np.indices((2 * n + 1, 2 * n + 1)) from sklearn.feature_extraction import image
mask = (x - n) ** 2 + (y - n) ** 2 <= n**2 from sklearn.cluster import spectral_clustering
struct[mask] = 1
return struct.astype(bool) l = 100
x, y = np.indices((l, l))

def granulometry(data, sizes=None): center1 = (28, 24)

s = max(data.shape) center2 = (40, 50)
if sizes is None: center3 = (67, 58)
sizes = range(1, s / 2, 2) center4 = (24, 70)
granulo = [
sp.ndimage.binary_opening(data, structure=disk_structure(n)).sum() radius1, radius2, radius3, radius4 = 16, 14, 15, 14
for n in sizes
] circle1 = (x - center1[0]) ** 2 + (y - center1[1]) ** 2 < radius1**2
return granulo circle2 = (x - center2[0]) ** 2 + (y - center2[1]) ** 2 < radius2**2
circle3 = (x - center3[0]) ** 2 + (y - center3[1]) ** 2 < radius3**2
circle4 = (x - center4[0]) ** 2 + (y - center4[1]) ** 2 < radius4**2
rng = np.random.default_rng(27446968)
n = 10 4 circles
l = 256
im = np.zeros((l, l)) img = circle1 + circle2 + circle3 + circle4
points = l * rng.random((2, n**2)) mask = img.astype(bool)
(continues on next page) (continues on next page)

12.8. Examples for the image processing chapter 428 12.8. Examples for the image processing chapter 429
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

ã See also
img = img.astype(float)
More on image-processing:
rng = np.random.default_rng(27446968)
img += 1 + 0.2 * rng.normal(size=img.shape) • The chapter on Scikit-image
• Other, more powerful and complete modules: OpenCV (Python bindings), CellProfiler, ITK
# Convert the image into a graph with the value of the gradient on the with Python bindings
# edges.
graph = image.img_to_graph(img, mask=mask)

# Take a decreasing function of the gradient: we take it weakly

# dependent from the gradient the segmentation is close to a voronoi
graph.data = np.exp(-graph.data / graph.data.std())

# Force the solver to be arpack, since amg is numerically

# unstable on this example
labels = spectral_clustering(graph, n_clusters=4)
label_im = -np.ones(mask.shape)
label_im[mask] = labels

plt.figure(figsize=(6, 3))
plt.subplot(121)
plt.imshow(img, cmap="nipy_spectral", interpolation="nearest")
plt.axis("off")
plt.subplot(122)
plt.imshow(label_im, cmap="nipy_spectral", interpolation="nearest")
plt.axis("off")

plt.subplots_adjust(wspace=0, hspace=0.0, top=0.99, bottom=0.01, left=0.01, right=0.

˓→99)

plt.show()

Total running time of the script: (0 minutes 0.257 seconds)

12.8. Examples for the image processing chapter 430 12.8. Examples for the image processing chapter 431
 Scientific Python Lectures, Edition 2025.1rc0.dev0

• Practical Methods of Optimization by Fletcher: good at hand-waving explanations.

Chapters contents

13
• Knowing your problem
– Convex versus non-convex optimization
– Smooth and non-smooth problems
– Noisy versus exact cost functions
CHAPTER
– Constraints
• A review of the different optimizers
– Getting started: 1D optimization
– Gradient based methods
– Newton and quasi-newton methods
• Full code examples
Mathematical optimization: finding • Examples for the mathematical optimization chapter

minima of functions
– Gradient-less methods
– Global optimizers
• Practical guide to optimization with SciPy
– Choosing a method
– Making your optimizer faster
– Computing gradients
Authors: Gaël Varoquaux – Synthetic exercises
Mathematical optimization deals with the problem of finding numerically minimums (or maximums or • Special case: non-linear least-squares
zeros) of a function. In this context, the function is called cost function, or objective function, or energy.
– Minimizing the norm of a vector function
Here, we are interested in using scipy.optimize for black-box optimization: we do not rely on the
– Curve fitting
mathematical expression of the function that we are optimizing. Note that this expression can often be
used for more efficient, non black-box, optimization. • Optimization with constraints
– Box bounds
Prerequisites – General constraints
• NumPy • Full code examples
• SciPy • Examples for the mathematical optimization chapter
• Matplotlib

ã See also 13.1 Knowing your problem

References Not all optimization problems are equal. Knowing your problem enables you to choose the right tool.
Mathematical optimization is very . . . mathematical. If you want performance, it really pays to read
the books: Dimensionality of the problem
• Convex Optimization by Boyd and Vandenberghe (pdf available free online). The scale of an optimization problem is pretty much set by the dimensionality of the problem, i.e.
• Numerical Optimization, by Nocedal and Wright. Detailed reference on gradient descent meth- the number of scalar variables on which the search is performed.
ods.

432 13.1. Knowing your problem 433

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

13.1.1 Convex versus non-convex optimization 13.1.3 Noisy versus exact cost functions

Noisy (blue) and non-noisy (green) functions

A convex function: A non-convex function
• f is above all its tangents.
• equivalently, for two point A, B, f(C) lies Noisy gradients
below the segment [f(A), f(B])], if A < C <
B Many optimization methods rely on gradients of the objective function. If the gradient function is not
given, they are computed numerically, which induces errors. In such situation, even if the objective
function is not noisy, a gradient-based optimization may be a noisy optimization.
Optimizing convex functions is easy. Optimizing non-convex functions can be very hard.
13.1.4 Constraints
ò Note

It can be proven that for a convex function a local minimum is also a global minimum. Then, in
some sense, the minimum is unique.

13.1.2 Smooth and non-smooth problems

Optimizations under constraints

Here:
−1 < 𝑥1 < 1
−1 < 𝑥2 < 1

13.2 A review of the different optimizers

13.2.1 Getting started: 1D optimization
A smooth function: A non-smooth function Let’s get started by finding the minimum of the scalar function 𝑓 (𝑥) = exp[(𝑥−0.5)2 ]. scipy.optimize.
The gradient is defined everywhere, and is a continuous function minimize_scalar() uses Brent’s method to find the minimum of a function:

>>> import numpy as np

Optimizing smooth functions is easier (true in the context of black-box optimization, otherwise Lin- >>> import scipy as sp
ear Programming is an example of methods which deal very efficiently with piece-wise linear functions). >>> def f(x):
(continues on next page)

13.1. Knowing your problem 434 13.2. A review of the different optimizers 435
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

scipy.optimize.minimize_scalar() can also be used for optimization constrained to an interval
... return -np.exp(-(x - 0.5)**2) using the parameter bounds.
>>> result = sp.optimize.minimize_scalar(f)
>>> result.success # check if solver was successful
True 13.2.2 Gradient based methods
>>> x_min = result.x
>>> x_min Some intuitions about gradient descent
np.float64(0.50...)
>>> x_min - 0.5 Here we focus on intuitions, not code. Code will follow.
np.float64(5.8...e-09) Gradient descent basically consists in taking small steps in the direction of the gradient, that is the
direction of the steepest descent.
Table 1: Brent’s method on a quadratic function: it converges
Table 3: Fixed step gradient descent
in 3 iterations, as the quadratic approximation is then exact.

A well-conditioned
quadratic function.

Table 2: Brent’s method on a non-convex function: note that

the fact that the optimizer avoided the local minimum is a matter
of luck.
An ill-conditioned
quadratic function.
The core problem of gradient-
methods on ill-conditioned
problems is that the gradient
tends not to point in the direc-
tion of the minimum.

We can see that very anisotropic (ill-conditioned) functions are harder to optimize.

Take home message: conditioning number and preconditioning

If you know natural scaling for your variables, prescale them so that they behave similarly. This is
related to preconditioning.

Also, it clearly can be advantageous to take bigger steps. This is done in gradient descent code using a
ò Note
line search.
You can use different solvers using the parameter method.

ò Note

13.2. A review of the different optimizers 436 13.2. A review of the different optimizers 437
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Table 4: Adaptive step gradient descent Table 5: Conjugate gradient descent

A well-conditioned quadratic An ill-conditioned non-

function. quadratic function.

An ill-conditioned quadratic An ill-conditioned very non-

function. quadratic function.

SciPy provides scipy.optimize.minimize() to find the minimum of scalar functions of one or more
variables. The simple conjugate gradient method can be used by setting the parameter method to CG

>>> def f(x): # The rosenbrock function

... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> sp.optimize.minimize(f, [2, -1], method="CG")
message: Optimization terminated successfully.
success: True
An ill-conditioned non- status: 0
quadratic function. fun: 1.650...e-11
x: [ 1.000e+00 1.000e+00]
nit: 13
jac: [-6.15...e-06 2.53...e-07]
nfev: 81
njev: 27

Gradient methods need the Jacobian (gradient) of the function. They can compute it numerically, but
will perform better if you can pass them the gradient:
An ill-conditioned very non- >>> def jacobian(x):
quadratic function. ... return np.array((-2*.5*(1 - x[0]) - 4*x[0]*(x[1] - x[0]**2), 2*(x[1] -␣
˓→x[0]**2)))

The more a function looks like a quadratic function (elliptic iso-curves), the easier it is to optimize. >>> sp.optimize.minimize(f, [2, 1], method="CG", jac=jacobian)
message: Optimization terminated successfully.
success: True
Conjugate gradient descent
status: 0
The gradient descent algorithms above are toys not to be used on real problems. fun: 2.95786...e-14
x: [ 1.000e+00 1.000e+00]
As can be seen from the above experiments, one of the problems of the simple gradient descent algorithms,
nit: 8
is that it tends to oscillate across a valley, each time following the direction of the gradient, that makes
jac: [ 7.183e-07 -2.990e-07]
it cross the valley. The conjugate gradient solves this problem by adding a friction term: each step
nfev: 16
depends on the two last values of the gradient and sharp turns are reduced.
njev: 16

13.2. A review of the different optimizers 438 13.2. A review of the different optimizers 439
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Note that the function has only been evaluated 27 times, compared to 108 without the gradient. (continued from previous page)
>>> def jacobian(x):
13.2.3 Newton and quasi-newton methods ... return np.array((-2*.5*(1 - x[0]) - 4*x[0]*(x[1] - x[0]**2), 2*(x[1] -␣
˓→x[0]**2)))
Newton methods: using the Hessian (2nd differential) >>> sp.optimize.minimize(f, [2,-1], method="Newton-CG", jac=jacobian)
Newton methods use a local quadratic approximation to compute the jump direction. For this purpose, message: Optimization terminated successfully.
they rely on the 2 first derivative of the function: the gradient and the Hessian. success: True
status: 0
fun: 1.5601357400786612e-15
x: [ 1.000e+00 1.000e+00]
nit: 10
jac: [ 1.058e-07 -7.483e-08]
nfev: 11
njev: 33
nhev: 0

Note that compared to a conjugate gradient (above), Newton’s method has required less function evalua-
tions, but more gradient evaluations, as it uses it to approximate the Hessian. Let’s compute the Hessian
An ill-conditioned and pass it to the algorithm:
quadratic function:
Note that, as the quadratic ap- >>> def hessian(x): # Computed with sympy
proximation is exact, the New- ... return np.array(((1 - 4*x[1] + 12*x[0]**2, -4*x[0]), (-4*x[0], 2)))
ton method is blazing fast >>> sp.optimize.minimize(f, [2,-1], method="Newton-CG", jac=jacobian, hess=hessian)
message: Optimization terminated successfully.
success: True
status: 0
fun: 1.6277298383706738e-15
x: [ 1.000e+00 1.000e+00]
nit: 10
jac: [ 1.110e-07 -7.781e-08]
nfev: 11
njev: 11
An ill-conditioned non- nhev: 10
quadratic function:
Here we are optimizing a Gaus-
sian, which is always below its ò Note
quadratic approximation. As a
result, the Newton method over- At very high-dimension, the inversion of the Hessian can be costly and unstable (large scale > 250).
shoots and leads to oscillations.

ò Note

Newton optimizers should not to be confused with Newton’s root finding method, based on the same
principles, scipy.optimize.newton().

Quasi-Newton methods: approximating the Hessian on the fly

An ill-conditioned very non- BFGS: BFGS (Broyden-Fletcher-Goldfarb-Shanno algorithm) refines at each step an approximation of
quadratic function: the Hessian.

In SciPy, you can use the Newton method by setting method to Newton-CG in scipy.optimize.
minimize(). Here, CG refers to the fact that an internal inversion of the Hessian is performed by
conjugate gradient

>>> def f(x): # The rosenbrock function

... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
(continues on next page)

13.2. A review of the different optimizers 440 13.2. A review of the different optimizers 441
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

13.3 Full code examples

13.4 Examples for the mathematical optimization chapter
13.4.1 Noisy optimization problem
Draws a figure explaining noisy vs non-noisy optimization

import numpy as np
import matplotlib.pyplot as plt

rng = np.random.default_rng(27446968)

x = np.linspace(-5, 5, 101)
x_ = np.linspace(-5, 5, 31) •

import numpy as np
def f(x): import matplotlib.pyplot as plt
return -np.exp(-(x**2))
x = np.linspace(-1.5, 1.5, 101)

# A smooth function # A smooth function

plt.figure(1, figsize=(3, 2.5)) plt.figure(1, figsize=(3, 2.5))
plt.clf() plt.clf()

plt.plot(x_, f(x_) + 0.2 * np.random.normal(size=31), linewidth=2) plt.plot(x, np.sqrt(0.2 + x**2), linewidth=2)

plt.plot(x, f(x), linewidth=2) plt.text(-1, 0, "$f$", size=20)

plt.ylim(ymin=-1.3) plt.ylim(ymin=-0.2)
plt.axis("off") plt.axis("off")
plt.tight_layout() plt.tight_layout()
plt.show()
# A non-smooth function
Total running time of the script: (0 minutes 0.015 seconds) plt.figure(2, figsize=(3, 2.5))
plt.clf()
plt.plot(x, np.abs(x), linewidth=2)
plt.text(-1, 0, "$f$", size=20)
13.4.2 Smooth vs non-smooth
plt.ylim(ymin=-0.2)
Draws a figure to explain smooth versus non smooth optimization. plt.axis("off")
plt.tight_layout()
plt.show()

13.3. Full code examples 442 13.4. Examples for the mathematical optimization chapter 443
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Total running time of the script: (0 minutes 0.037 seconds) (continued from previous page)
t = np.linspace(0, 3, 1000)

plt.figure(1)
13.4.3 Curve fitting plt.clf()
A curve fitting example plt.plot(x, y, "bx")
plt.plot(t, f(t, *params), "r-")
plt.show()

Total running time of the script: (0 minutes 0.054 seconds)

13.4.4 Convex function

A figure showing the definition of a convex function

import numpy as np
import scipy as sp
import matplotlib.pyplot as plt

rng = np.random.default_rng(27446968)

# Our test function

def f(t, omega, phi):
return np.cos(omega * t + phi)
•

# Our x and y data import numpy as np

x = np.linspace(0, 3, 50) import matplotlib.pyplot as plt
y = f(x, 1.5, 1) + 0.1 * np.random.normal(size=50)
x = np.linspace(-1, 2)
# Fit the model: the parameters omega and phi can be found in the
# `params` vector plt.figure(1, figsize=(3, 2.5))
params, params_cov = sp.optimize.curve_fit(f, x, y) plt.clf()

# plot the data and the fitted curve # A convex function

(continues on next page) (continues on next page)

13.4. Examples for the mathematical optimization chapter 444 13.4. Examples for the mathematical optimization chapter 445
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

plt.plot(x, x**2, linewidth=2)
plt.text(-0.7, -(0.6**2), "$f$", size=20) # The gradient of g. We won't use it here for the optimization.
def g_prime(x):
# The tangent in one point r = np.sqrt(x[0] ** 2 + x[1] ** 2)
plt.plot(x, 2 * x - 1) return 2 / r**3 * g(x) * x / r
plt.plot(1, 1, "k+")
plt.text(0.3, -0.75, "Tangent to $f$", size=15)
plt.text(1, 1 - 0.5, "C", size=15) result = sp.optimize.minimize(g, [1, 1], method="Powell", tol=1e-10)
x_min = result.x
# Convexity as barycenter
plt.plot([0.35, 1.85], [0.35**2, 1.85**2]) Some pretty plotting
plt.plot([0.35, 1.85], [0.35**2, 1.85**2], "k+")
plt.text(0.35 - 0.2, 0.35**2 + 0.1, "A", size=15) plt.figure(0)
plt.text(1.85 - 0.2, 1.85**2, "B", size=15) plt.clf()
t = np.linspace(-1.1, 1.1, 100)
plt.ylim(ymin=-1) plt.plot(t, f([0, t]))
plt.axis("off")
plt.tight_layout() plt.figure(1)
plt.clf()
# Convexity as barycenter X, Y = np.mgrid[-1.5:1.5:100j, -1.1:1.1:100j] # type: ignore[misc]
plt.figure(2, figsize=(3, 2.5)) plt.imshow(f([X, Y]).T, cmap="gray_r", extent=(-1.5, 1.5, -1.1, 1.1), origin="lower")
plt.clf() plt.contour(X, Y, f([X, Y]), cmap="gnuplot")
plt.plot(x, x**2 + np.exp(-5 * (x - 0.5) ** 2), linewidth=2)
plt.text(-0.7, -(0.6**2), "$f$", size=20) # Plot the gradient
dX, dY = g_prime([0.1 * X[::5, ::5], Y[::5, ::5]])
plt.ylim(ymin=-1) # Adjust for our preconditioning
plt.axis("off") dX *= 0.1
plt.tight_layout() plt.quiver(X[::5, ::5], Y[::5, ::5], dX, dY, color=".5")
plt.show()
# Plot our solution
Total running time of the script: (0 minutes 0.046 seconds) plt.plot(x_min[0], x_min[1], "r+", markersize=15)

plt.show()

13.4.5 Finding a minimum in a flat neighborhood

An exercise of finding minimum. This exercise is hard because the function is very flat around the
minimum (all its derivatives are zero). Thus gradient information is unreliable.
The function admits a minimum in [0, 0]. The challenge is to get within 1e-7 of this minimum, starting
at x0 = [1, 1].
The solution that we adopt here is to give up on using gradient or information based on local differences,
and to rely on the Powell algorithm. With 162 function evaluations, we get to 1e-8 of the solution.

import numpy as np
import scipy as sp
import matplotlib.pyplot as plt

def f(x):
return np.exp(-1 / (0.01 * x[0] ** 2 + x[1] ** 2))

# A well-conditionned version of f:
def g(x):
return f([10 * x[0], x[1]])

(continues on next page)

13.4. Examples for the mathematical optimization chapter 446 13.4. Examples for the mathematical optimization chapter 447
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

13.4.6 Optimization with constraints

An example showing how to do optimization with general constraints using SLSQP and cobyla.

import numpy as np
import matplotlib.pyplot as plt
import scipy as sp

x, y = np.mgrid[-2.03:4.2:0.04, -1.6:3.2:0.04] # type: ignore[misc]

x = x.T
y = y.T
•
plt.figure(1, figsize=(3, 2.5))
plt.clf()
plt.axes((0, 0, 1, 1))

contours = plt.contour(
np.sqrt((x - 3) ** 2 + (y - 2) ** 2),
extent=[-2.03, 4.2, -1.6, 3.2],
cmap="gnuplot",
)
plt.clabel(contours, inline=1, fmt="%1.1f ", fontsize=14)
plt.plot([-1.5, 0, 1.5, 0, -1.5], [0, 1.5, 0, -1.5, 0], "k", linewidth=2)
plt.fill_between([-1.5, 0, 1.5], [0, -1.5, 0], [0, 1.5, 0], color=".8")
plt.axvline(0, color="k")
plt.axhline(0, color="k")

plt.text(-0.9, 2.8, "$x_2$", size=20)

plt.text(3.6, -0.6, "$x_1$", size=20)
plt.axis("tight")
plt.axis("off")

# And now plot the optimization path

accumulator = []

def f(x):
# Store the list of function calls
• accumulator.append(x)
return np.sqrt((x[0] - 3) ** 2 + (x[1] - 2) ** 2)
Total running time of the script: (0 minutes 0.117 seconds)

(continues on next page)

13.4. Examples for the mathematical optimization chapter 448 13.4. Examples for the mathematical optimization chapter 449
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

def constraint(x):
return np.atleast_1d(1.5 - np.sum(np.abs(x)))

sp.optimize.minimize(
f, np.array([0, 0]), method="SLSQP", constraints={"fun": constraint, "type": "ineq
˓→"}

accumulated = np.array(accumulator)
plt.plot(accumulated[:, 0], accumulated[:, 1])

plt.show()
•
Total running time of the script: (0 minutes 0.050 seconds)

13.4.7 Brent’s method

Illustration of 1D optimization: Brent’s method

Converged at 6
Converged at 23
•

import numpy as np
import matplotlib.pyplot as plt
import scipy as sp

x = np.linspace(-1, 3, 100)
x_0 = np.exp(-1)

def f(x):
return (x - x_0) ** 2 + epsilon * np.exp(-5 * (x - 0.5 - x_0) ** 2)

for epsilon in (0, 1):

• plt.figure(figsize=(3, 2.5))
plt.axes((0, 0, 1, 1))
(continues on next page)

13.4. Examples for the mathematical optimization chapter 450 13.4. Examples for the mathematical optimization chapter 451
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

# A convex function
plt.plot(x, f(x), linewidth=2)

# Apply brent method. To have access to the iteration, do this in an

# artificial way: allow the algorithm to iter only once
all_x = []
all_y = []
for iter in range(30):
result = sp.optimize.minimize_scalar(
f,
bracket=(-5, 2.9, 4.5),
method="Brent",
options={"maxiter": iter}, •
tol=np.finfo(1.0).eps,
)
if result.success:
print("Converged at ", iter)
break

this_x = result.x
all_x.append(this_x)
all_y.append(f(this_x))
if iter < 6:
plt.text(
this_x - 0.05 * np.sign(this_x) - 0.05,
f(this_x) + 1.2 * (0.3 - iter % 2),
str(iter + 1),
size=12,
) •
import numpy as np
plt.plot(all_x[:10], all_y[:10], "k+", markersize=12, markeredgewidth=2) import matplotlib.pyplot as plt
import scipy as sp
plt.plot(all_x[-1], all_y[-1], "rx", markersize=12)
plt.axis("off") x, y = np.mgrid[-2.9:5.8:0.05, -2.5:5:0.05] # type: ignore[misc]
plt.ylim(ymin=-1, ymax=8) x = x.T
y = y.T
plt.figure(figsize=(4, 3))
plt.semilogy(np.abs(all_y - all_y[-1]), linewidth=2) for i in (1, 2):
plt.ylabel("Error on f(x)") # Create 2 figure: only the second one will have the optimization
plt.xlabel("Iteration") # path
plt.tight_layout() plt.figure(i, figsize=(3, 2.5))
plt.clf()
plt.show() plt.axes((0, 0, 1, 1))

Total running time of the script: (0 minutes 0.233 seconds) contours = plt.contour(
np.sqrt((x - 3) ** 2 + (y - 2) ** 2),
extent=[-3, 6, -2.5, 5],
cmap="gnuplot",
13.4.8 Constraint optimization: visualizing the geometry )
A small figure explaining optimization with constraints plt.clabel(contours, inline=1, fmt="%1.1f ", fontsize=14)
plt.plot(
[-1.5, -1.5, 1.5, 1.5, -1.5], [-1.5, 1.5, 1.5, -1.5, -1.5], "k", linewidth=2
)
plt.fill_between([-1.5, 1.5], [-1.5, -1.5], [1.5, 1.5], color=".8")
plt.axvline(0, color="k")
(continues on next page)

13.4. Examples for the mathematical optimization chapter 452 13.4. Examples for the mathematical optimization chapter 453
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

plt.axhline(0, color="k")
def f_prime(x):
plt.text(-0.9, 4.4, "$x_2$", size=20) return 2 * K.T @ K @ (x - 1) + 4 * np.sum(x**2) * x
plt.text(5.6, -0.6, "$x_1$", size=20)
plt.axis("equal")
plt.axis("off") def hessian(x):
H = 2 * K.T @ K + 4 * 2 * x * x[:, np.newaxis]
# And now plot the optimization path return H + 4 * np.eye(H.shape[0]) * np.sum(x**2)
accumulator = []
Some pretty plotting

def f(x): plt.figure(1)

# Store the list of function calls plt.clf()
accumulator.append(x) Z = X, Y = np.mgrid[-1.5:1.5:100j, -1.1:1.1:100j] # type: ignore[misc]
return np.sqrt((x[0] - 3) ** 2 + (x[1] - 2) ** 2) # Complete in the additional dimensions with zeros
Z = np.reshape(Z, (2, -1)).copy()
Z.resize((100, Z.shape[-1]))
# We don't use the gradient, as with the gradient, L-BFGS is too fast, Z = np.apply_along_axis(f, 0, Z)
# and finds the optimum without showing us a pretty path Z = np.reshape(Z, X.shape)
def f_prime(x): plt.imshow(Z.T, cmap="gray_r", extent=(-1.5, 1.5, -1.1, 1.1), origin="lower")
r = np.sqrt((x[0] - 3) ** 2 + (x[0] - 2) ** 2) plt.contour(X, Y, Z, cmap="gnuplot")
return np.array(((x[0] - 3) / r, (x[0] - 2) / r))
# A reference but slow solution:
t0 = time.time()
sp.optimize.minimize( x_ref = sp.optimize.minimize(f, K[0], method="Powell").x
f, np.array([0, 0]), method="L-BFGS-B", bounds=((-1.5, 1.5), (-1.5, 1.5)) print(f" Powell: time { time.time() - t0: .2f} s")
) f_ref = f(x_ref)

accumulated = np.array(accumulator) # Compare different approaches

plt.plot(accumulated[:, 0], accumulated[:, 1]) t0 = time.time()
x_bfgs = sp.optimize.minimize(f, K[0], method="BFGS").x
plt.show() print(
f" BFGS: time { time.time() - t0: .2f} s, x error { np.sqrt(np.sum((x_bfgs -␣
˓→x_ref) ** 2)): .2f} , f error { f(x_bfgs) - f_ref: .2f} "
Total running time of the script: (0 minutes 0.084 seconds)
)

t0 = time.time()
13.4.9 Alternating optimization x_l_bfgs = sp.optimize.minimize(f, K[0], method="L-BFGS-B").x
print(
The challenge here is that Hessian of the problem is a very ill-conditioned matrix. This can easily be f" L-BFGS: time { time.time() - t0: .2f} s, x error { np.sqrt(np.sum((x_l_bfgs -
seen, as the Hessian of the first term in simply 2 * K.T @ K. Thus the conditioning of the problem can ˓→ x_ref) ** 2)): .2f} , f error { f(x_l_bfgs) - f_ref: .2f} "
be judged from looking at the conditioning of K. )
import time

import numpy as np t0 = time.time()

import scipy as sp x_bfgs = sp.optimize.minimize(f, K[0], jac=f_prime, method="BFGS").x
import matplotlib.pyplot as plt print(
f" BFGS w f': time { time.time() - t0: .2f} s, x error { np.sqrt(np.sum((x_bfgs -␣
˓→x_ref) ** 2)): .2f} , f error { f(x_bfgs) - f_ref: .2f} "
rng = np.random.default_rng(27446968)
)
K = rng.normal(size=(100, 100))
t0 = time.time()
x_l_bfgs = sp.optimize.minimize(f, K[0], jac=f_prime, method="L-BFGS-B").x
def f(x): print(
return np.sum((K @ (x - 1)) ** 2) + np.sum(x**2) ** 2 f"L-BFGS w f': time { time.time() - t0: .2f} s, x error { np.sqrt(np.sum((x_l_bfgs -
˓→ x_ref) ** 2)): .2f} , f error { f(x_l_bfgs) - f_ref: .2f} "

(continues on next page) (continues on next page)

13.4. Examples for the mathematical optimization chapter 454 13.4. Examples for the mathematical optimization chapter 455
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) 13.4.10 Plotting the comparison of optimizers

) Plots the results from the comparison of optimizers.
t0 = time.time()
x_newton = sp.optimize.minimize(
f, K[0], jac=f_prime, hess=hessian, method="Newton-CG"
).x
print(
f" Newton: time { time.time() - t0: .2f} s, x error { np.sqrt(np.sum((x_newton -
˓→ x_ref) ** 2)): .2f} , f error { f(x_newton) - f_ref: .2f} "

plt.show()

import pickle
import sys

import numpy as np

import matplotlib
import matplotlib.pyplot as plt

results = pickle.load(
open(f"helper/compare_optimizers_py{ sys.version_info[0]} .pkl", "rb")
)
n_methods = len(list(results.values())[0]["Rosenbrock "])
n_dims = len(results)

symbols = "o>*Ds"

plt.figure(1, figsize=(10, 4))

plt.clf()

nipy_spectral = matplotlib.colormaps["nipy_spectral"]
colors = nipy_spectral(np.linspace(0, 1, n_dims))[:, :3]

method_names = list(list(results.values())[0]["Rosenbrock "].keys())

method_names.sort(key=lambda x: x[::-1], reverse=True)
Powell: time 0.14s
BFGS: time 0.60s, x error 0.02, f error -0.03 for n_dim_index, ((n_dim, n_dim_bench), color) in enumerate(
L-BFGS: time 0.13s, x error 0.02, f error -0.03 zip(sorted(results.items()), colors, strict=True)
BFGS w f': time 0.15s, x error 0.02, f error -0.03 ):
L-BFGS w f': time 0.08s, x error 0.02, f error -0.03 for (cost_name, cost_bench), symbol in zip(
Newton: time 0.01s, x error 0.02, f error -0.03 sorted(n_dim_bench.items()), symbols, strict=True
):
Total running time of the script: (0 minutes 1.341 seconds) for (
method_index,
method_name,
) in enumerate(method_names):
this_bench = cost_bench[method_name]
bench = np.mean(this_bench)
plt.semilogy(
(continues on next page)

13.4. Examples for the mathematical optimization chapter 456 13.4. Examples for the mathematical optimization chapter 457
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

[
method_index + 0.1 * n_dim_index, plt.tight_layout()
], plt.show()
[
bench, Total running time of the script: (0 minutes 0.571 seconds)
],
marker=symbol,
color=color,
) 13.4.11 Gradient descent
An example demoing gradient descent by creating figures that trace the evolution of the optimizer.
# Create a legend for the problem type
for cost_name, symbol in zip(sorted(n_dim_bench.keys()), symbols, strict=True): import numpy as np
plt.semilogy( import matplotlib.pyplot as plt
[ import scipy as sp
-10,
], import collections
[ import sys
0, import os
],
symbol, sys.path.append(os.path.abspath("helper"))
color=".5", from cost_functions import (
label=cost_name, mk_quad,
) mk_gauss,
rosenbrock,
plt.xticks(np.arange(n_methods), method_names, size=11) rosenbrock_prime,
plt.xlim(-0.2, n_methods - 0.5) rosenbrock_hessian,
plt.legend(loc="best", numpoints=1, handletextpad=0, prop={"size": 12}, frameon=False) LoggingFunction,
plt.ylabel("# function calls (a.u.)") CountingFunction,
)
# Create a second legend for the problem dimensionality
plt.twinx() x_min, x_max = -1, 2
y_min, y_max = 2.25 / 3 * x_min - 0.2, 2.25 / 3 * x_max - 0.2
for n_dim, color in zip(sorted(results.keys()), colors, strict=True):
plt.plot( A formatter to print values on contours
[
-10, def super_fmt(value):
], if value > 1:
[ if np.abs(int(value) - value) < 0.1:
0, out = f"$10^{{{ int(value): d}}} $"
], else:
"o", out = f"$10^{{{ value: .1f}}} $"
color=color, else:
label=f"# dim: { n_dim} ", value = np.exp(value - 0.01)
) if value > 0.1:
plt.legend( out = f"{ value: 1.1f} "
loc=(0.47, 0.07), elif value > 0.01:
numpoints=1, out = f"{ value: .2f} "
handletextpad=0, else:
prop={"size": 12}, out = f"{ value: .2e} "
frameon=False, return out
ncol=2,
) A gradient descent algorithm do not use: its a toy, use scipy’s optimize.fmin_cg
plt.xlim(-0.2, n_methods - 0.5)
def gradient_descent(x0, f, f_prime, hessian=None, adaptative=False):
plt.xticks(np.arange(n_methods), method_names) x_i, y_i = x0
plt.yticks(()) all_x_i = []
(continues on next page) (continues on next page)

13.4. Examples for the mathematical optimization chapter 458 13.4. Examples for the mathematical optimization chapter 459
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

all_y_i = []
all_f_i = [] def store(X):
x, y = X
for i in range(1, 100): all_x_i.append(x)
all_x_i.append(x_i) all_y_i.append(y)
all_y_i.append(y_i) all_f_i.append(f(X))
all_f_i.append(f([x_i, y_i]))
dx_i, dy_i = f_prime(np.asarray([x_i, y_i])) sp.optimize.minimize(
if adaptative: f,
# Compute a step size using a line_search to satisfy the Wolf x0,
# conditions method="Newton-CG",
step = sp.optimize.line_search( jac=f_prime,
f, hess=hessian,
f_prime, callback=store,
np.r_[x_i, y_i], options={"xtol": 1e-12},
-np.r_[dx_i, dy_i], )
np.r_[dx_i, dy_i], return all_x_i, all_y_i, all_f_i
c2=0.05,
)
step = step[0] def bfgs(x0, f, f_prime, hessian=None):
if step is None: all_x_i = [x0[0]]
step = 0 all_y_i = [x0[1]]
else: all_f_i = [f(x0)]
step = 1
x_i += -step * dx_i def store(X):
y_i += -step * dy_i x, y = X
if np.abs(all_f_i[-1]) < 1e-16: all_x_i.append(x)
break all_y_i.append(y)
return all_x_i, all_y_i, all_f_i all_f_i.append(f(X))

sp.optimize.minimize(
def gradient_descent_adaptative(x0, f, f_prime, hessian=None): f, x0, method="BFGS", jac=f_prime, callback=store, options={"gtol": 1e-12}
return gradient_descent(x0, f, f_prime, adaptative=True) )
return all_x_i, all_y_i, all_f_i

def conjugate_gradient(x0, f, f_prime, hessian=None):

all_x_i = [x0[0]] def powell(x0, f, f_prime, hessian=None):
all_y_i = [x0[1]] all_x_i = [x0[0]]
all_f_i = [f(x0)] all_y_i = [x0[1]]
all_f_i = [f(x0)]
def store(X):
x, y = X def store(X):
all_x_i.append(x) x, y = X
all_y_i.append(y) all_x_i.append(x)
all_f_i.append(f(X)) all_y_i.append(y)
all_f_i.append(f(X))
sp.optimize.minimize(
f, x0, jac=f_prime, method="CG", callback=store, options={"gtol": 1e-12} sp.optimize.minimize(
) f, x0, method="Powell", callback=store, options={"ftol": 1e-12}
return all_x_i, all_y_i, all_f_i )
return all_x_i, all_y_i, all_f_i

def newton_cg(x0, f, f_prime, hessian):

all_x_i = [x0[0]] def nelder_mead(x0, f, f_prime, hessian=None):
all_y_i = [x0[1]] all_x_i = [x0[0]]
all_f_i = [f(x0)] all_y_i = [x0[1]]
(continues on next page) (continues on next page)

13.4. Examples for the mathematical optimization chapter 460 13.4. Examples for the mathematical optimization chapter 461
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

all_f_i = [f(x0)] y_min *= 1.2
y_max *= 1.2
def store(X): x, y = np.mgrid[x_min:x_max:100j, y_min:y_max:100j]
x, y = X x = x.T
all_x_i.append(x) y = y.T
all_y_i.append(y)
all_f_i.append(f(X)) plt.figure(index, figsize=(3, 2.5))
plt.clf()
sp.optimize.minimize( plt.axes([0, 0, 1, 1])
f, x0, method="Nelder-Mead", callback=store, options={"ftol": 1e-12}
) X = np.concatenate((x[np.newaxis, ...], y[np.newaxis, ...]), axis=0)
return all_x_i, all_y_i, all_f_i z = np.apply_along_axis(f, 0, X)
log_z = np.log(z + 0.01)
Run different optimizers on these problems plt.imshow(
log_z,
levels = {} extent=[x_min, x_max, y_min, y_max],
cmap=plt.cm.gray_r,
for index, ((f, f_prime, hessian), optimizer) in enumerate( origin="lower",
( vmax=log_z.min() + 1.5 * np.ptp(log_z),
(mk_quad(0.7), gradient_descent), )
(mk_quad(0.7), gradient_descent_adaptative), contours = plt.contour(
(mk_quad(0.02), gradient_descent), log_z,
(mk_quad(0.02), gradient_descent_adaptative), levels=levels.get(f),
(mk_gauss(0.02), gradient_descent_adaptative), extent=[x_min, x_max, y_min, y_max],
( cmap=plt.cm.gnuplot,
(rosenbrock, rosenbrock_prime, rosenbrock_hessian), origin="lower",
gradient_descent_adaptative, )
), levels[f] = contours.levels
(mk_gauss(0.02), conjugate_gradient), plt.clabel(contours, inline=1, fmt=super_fmt, fontsize=14)
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), conjugate_gradient),
(mk_quad(0.02), newton_cg), plt.plot(all_x_i, all_y_i, "b-", linewidth=2)
(mk_gauss(0.02), newton_cg), plt.plot(all_x_i, all_y_i, "k+")
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), newton_cg),
(mk_quad(0.02), bfgs), plt.plot(logging_f.all_x_i, logging_f.all_y_i, "k.", markersize=2)
(mk_gauss(0.02), bfgs),
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), bfgs), plt.plot([0], [0], "rx", markersize=12)
(mk_quad(0.02), powell),
(mk_gauss(0.02), powell), plt.xticks(())
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), powell), plt.yticks(())
(mk_gauss(0.02), nelder_mead), plt.xlim(x_min, x_max)
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), nelder_mead), plt.ylim(y_min, y_max)
) plt.draw()
):
# Compute a gradient-descent plt.figure(index + 100, figsize=(4, 3))
x_i, y_i = 1.6, 1.1 plt.clf()
counting_f_prime = CountingFunction(f_prime) plt.semilogy(np.maximum(np.abs(all_f_i), 1e-30), linewidth=2, label="# iterations
counting_hessian = CountingFunction(hessian) ˓→")
logging_f = LoggingFunction(f, counter=counting_f_prime.counter) plt.ylabel("Error on f(x)")
all_x_i, all_y_i, all_f_i = optimizer( plt.semilogy(
np.array([x_i, y_i]), logging_f, counting_f_prime, hessian=counting_hessian logging_f.counts,
) np.maximum(np.abs(logging_f.all_f_i), 1e-30),
linewidth=2,
# Plot the contour plot color="g",
if not max(all_y_i) < y_max: label="# function calls",
x_min *= 1.2 )
x_max *= 1.2 plt.legend(
(continues on next page) (continues on next page)

13.4. Examples for the mathematical optimization chapter 462 13.4. Examples for the mathematical optimization chapter 463
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

loc="upper right",
frameon=True,
prop={"size": 11},
borderaxespad=0,
handlelength=1.5,
handletextpad=0.5,
)
plt.tight_layout()
plt.draw()

13.4. Examples for the mathematical optimization chapter 464 13.4. Examples for the mathematical optimization chapter 465
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• •

• •

• •

13.4. Examples for the mathematical optimization chapter 466 13.4. Examples for the mathematical optimization chapter 467
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• •

• •

• •

13.4. Examples for the mathematical optimization chapter 468 13.4. Examples for the mathematical optimization chapter 469
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

13.4. Examples for the mathematical optimization chapter 470 13.4. Examples for the mathematical optimization chapter 471
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• •

• •

• •

13.4. Examples for the mathematical optimization chapter 472 13.4. Examples for the mathematical optimization chapter 473
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• •

• •

• •

13.4. Examples for the mathematical optimization chapter 474 13.4. Examples for the mathematical optimization chapter 475
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

An ill-conditioned
quadratic function:
On a exactly quadratic function,
BFGS is not as fast as Newton’s
method, but still very fast.

An ill-conditioned non-
quadratic function:
Here BFGS does better than
Newton, as its empirical esti-
mate of the curvature is better
than that given by the Hessian.

/opt/hostedtoolcache/Python/3.12.9/x64/lib/python3.12/site-packages/scipy/optimize/_
˓→linesearch.py:312: LineSearchWarning: The line search algorithm did not converge

alpha_star, phi_star, old_fval, derphi_star = scalar_search_wolfe2(

/home/runner/work/scientific-python-lectures/scientific-python-lectures/advanced/ An ill-conditioned very non-
˓→mathematical_optimization/examples/plot_gradient_descent.py:70: LineSearchWarning:␣
quadratic function:
˓→The line search algorithm did not converge

step = sp.optimize.line_search(
>>> def f(x): # The rosenbrock function
/home/runner/work/scientific-python-lectures/scientific-python-lectures/advanced/
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
˓→mathematical_optimization/examples/plot_gradient_descent.py:234: RuntimeWarning:␣
>>> def jacobian(x):
˓→More than 20 figures have been opened. Figures created through the pyplot interface␣
... return np.array((-2*.5*(1 - x[0]) - 4*x[0]*(x[1] - x[0]**2), 2*(x[1] -␣
˓→(`matplotlib.pyplot.figure`) are retained until explicitly closed and may consume␣
˓→x[0]**2)))
˓→too much memory. (To control this warning, see the rcParam `figure.max_open_
>>> sp.optimize.minimize(f, [2, -1], method="BFGS", jac=jacobian)
˓→warning`). Consider using `matplotlib.pyplot.close()`.
message: Optimization terminated successfully.
plt.figure(index, figsize=(3, 2.5))
success: True
/home/runner/work/scientific-python-lectures/scientific-python-lectures/advanced/
status: 0
˓→mathematical_optimization/examples/plot_gradient_descent.py:179: OptimizeWarning:␣
fun: 2.630637192365927e-16
˓→Unknown solver options: ftol
x: [ 1.000e+00 1.000e+00]
sp.optimize.minimize(
nit: 8
jac: [ 6.709e-08 -3.222e-08]
Total running time of the script: (0 minutes 6.627 seconds)
(continues on next page)

13.4. Examples for the mathematical optimization chapter 476 13.4. Examples for the mathematical optimization chapter 477
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) Simplex method: the Nelder-Mead

hess_inv: [[ 9.999e-01 2.000e+00] The Nelder-Mead algorithms is a generalization of dichotomy approaches to high-dimensional spaces. The
[ 2.000e+00 4.499e+00]] algorithm works by refining a simplex, the generalization of intervals and triangles to high-dimensional
nfev: 10 spaces, to bracket the minimum.
njev: 10
Strong points: it is robust to noise, as it does not rely on computing gradients. Thus it can work
L-BFGS: Limited-memory BFGS Sits between BFGS and conjugate gradient: in very high dimensions on functions that are not locally smooth such as experimental data points, as long as they display a
(> 250) the Hessian matrix is too costly to compute and invert. L-BFGS keeps a low-rank version. In large-scale bell-shape behavior. However it is slower than gradient-based methods on smooth, non-noisy
addition, box bounds are also supported by L-BFGS-B: functions.

>>> def f(x): # The rosenbrock function

... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> def jacobian(x):
... return np.array((-2*.5*(1 - x[0]) - 4*x[0]*(x[1] - x[0]**2), 2*(x[1] -␣
˓→x[0]**2)))

>>> sp.optimize.minimize(f, [2, 2], method="L-BFGS-B", jac=jacobian)

message: CONVERGENCE: NORM OF PROJECTED GRADIENT <= PGTOL
success: True
status: 0
fun: 1.4417677473...e-15
An ill-conditioned non-
x: [ 1.000e+00 1.000e+00]
quadratic function:
nit: 16
jac: [ 1.023e-07 -2.593e-08]
nfev: 17
njev: 17
hess_inv: <2x2 LbfgsInvHessProduct with dtype=float64>

13.4.12 Gradient-less methods

A shooting method: the Powell algorithm
An ill-conditioned very non-
Almost a gradient approach quadratic function:

Using the Nelder-Mead solver in scipy.optimize.minimize():

>>> def f(x): # The rosenbrock function

... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> sp.optimize.minimize(f, [2, -1], method="Nelder-Mead")
message: Optimization terminated successfully.
success: True
status: 0
fun: 1.11527915993744e-10
An ill-conditioned x: [ 1.000e+00 1.000e+00]
quadratic function: nit: 58
Powell’s method isn’t too sensi- nfev: 111
tive to local ill-conditionning in final_simplex: (array([[ 1.000e+00, 1.000e+00],
low dimensions [ 1.000e+00, 1.000e+00],
[ 1.000e+00, 1.000e+00]]), array([ 1.115e-10, 1.537e-10, 4.
˓→988e-10]))

13.4.13 Global optimizers

If your problem does not admit a unique local minimum (which can be hard to test unless the function
is convex), and you do not have prior information to initialize the optimization close to the solution, you
may need a global optimizer.
An ill-conditioned very non-
quadratic function:

13.4. Examples for the mathematical optimization chapter 478 13.4. Examples for the mathematical optimization chapter 479
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Brute force: a grid search 13.5.2 Making your optimizer faster

scipy.optimize.brute() evaluates the function on a given grid of parameters and returns the param- • Choose the right method (see above), do compute analytically the gradient and Hessian, if you
eters corresponding to the minimum value. The parameters are specified with ranges given to numpy. can.
mgrid. By default, 20 steps are taken in each direction:
• Use preconditionning when possible.
>>> def f(x): # The rosenbrock function • Choose your initialization points wisely. For instance, if you are running many similar optimiza-
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2 tions, warm-restart one with the results of another.
>>> sp.optimize.brute(f, ((-1, 2), (-1, 2)))
array([1.0000..., 1.0000...]) • Relax the tolerance if you don’t need precision using the parameter tol.

13.5.3 Computing gradients

13.5 Practical guide to optimization with SciPy Computing gradients, and even more Hessians, is very tedious but worth the effort. Symbolic computa-
tion with Sympy may come in handy.
13.5.1 Choosing a method
All methods are exposed as the method argument of scipy.optimize.minimize(). . Warning

A very common source of optimization not converging well is human error in the computation of
the gradient. You can use scipy.optimize.check_grad() to check that your gradient is correct. It
returns the norm of the different between the gradient given, and a gradient computed numerically:
>>> sp.optimize.check_grad(f, jacobian, [2, -1])
np.float64(2.384185791015625e-07)

See also scipy.optimize.approx_fprime() to find your errors.

13.5.4 Synthetic exercises

Without knowledge of the gradient

• In general, prefer BFGS or L-BFGS, even if you have to approximate numerically
gradients. These are also the default if you omit the parameter method - depending
if the problem has constraints or bounds
• On well-conditioned problems, Powell and Nelder-Mead, both gradient-free
methods, work well in high dimension, but they collapse for ill-conditioned problems.
With knowledge of the gradient
Exercise: A simple (?) quadratic function
• BFGS or L-BFGS.
• Computational overhead of BFGS is larger than that L-BFGS, itself larger than Optimize the following function, using K[0] as a starting point:
that of conjugate gradient. On the other side, BFGS usually needs less function rng = np.random.default_rng(27446968)
evaluations than CG. Thus conjugate gradient method is better than BFGS at K = rng.normal(size=(100, 100))
optimizing computationally cheap functions.
With the Hessian def f(x):
return np.sum((K @ (x - 1))**2) + np.sum(x**2)**2
• If you can compute the Hessian, prefer the Newton method (Newton-CG or TCG).
Time your approach. Find the fastest approach. Why is BFGS not working well?
If you have noisy measurements
• Use Nelder-Mead or Powell.
Exercise: A locally flat minimum

Consider the function exp(-1/(.1*x**2 + y**2). This function admits a minimum in (0, 0). Starting
from an initialization at (1, 1), try to get within 1e-8 of this minimum point.

13.5. Practical guide to optimization with SciPy 480 13.5. Practical guide to optimization with SciPy 481
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

ò Note

leastsq is interesting compared to BFGS only if the dimensionality of the output vector is large, and
larger than the number of parameters to optimize.

. Warning

If the function is linear, this is a linear-algebra problem, and should be solved with scipy.linalg.
lstsq().

13.6.2 Curve fitting

Least square problems occur often when fitting a non-linear to data. While it is possible to construct our
optimization problem ourselves, SciPy provides a helper function for this purpose: scipy.optimize.
13.6 Special case: non-linear least-squares curve_fit():
13.6.1 Minimizing the norm of a vector function >>> def f(t, omega, phi):
Least square problems, minimizing the norm of a vector function, have a specific structure that can be ... return np.cos(omega * t + phi)
used in the Levenberg–Marquardt algorithm implemented in scipy.optimize.leastsq().
>>> x = np.linspace(0, 3, 50)
Lets try to minimize the norm of the following vectorial function: >>> rng = np.random.default_rng(27446968)
>>> y = f(x, 1.5, 1) + .1*rng.normal(size=50)
>>> def f(x):
... return np.arctan(x) - np.arctan(np.linspace(0, 1, len(x))) >>> sp.optimize.curve_fit(f, x, y)
(array([1.4812..., 0.9999...]), array([[ 0.0003..., -0.0004...],
>>> x0 = np.zeros(10) [-0.0004..., 0.0010...]]))
>>> sp.optimize.leastsq(f, x0)
(array([0. , 0.11111111, 0.22222222, 0.33333333, 0.44444444,
0.55555556, 0.66666667, 0.77777778, 0.88888889, 1. ]), ...)
Exercise
This took 67 function evaluations (check it with ‘full_output=True’). What if we compute the norm Do the same with omega = 3. What is the difficulty?
ourselves and use a good generic optimizer (BFGS):

>>> def g(x):

... return np.sum(f(x)**2) 13.7 Optimization with constraints
>>> result = sp.optimize.minimize(g, x0, method="BFGS")
>>> result.fun 13.7.1 Box bounds
np.float64(2.6940...e-11)
Box bounds correspond to limiting each of the individual parameters of the optimization. Note that
some problems that are not originally written as box bounds can be rewritten as such via change of vari-
BFGS needs more function calls, and gives a less precise result.
ables. Both scipy.optimize.minimize_scalar() and scipy.optimize.minimize() support bound
constraints with the parameter bounds:

13.6. Special case: non-linear least-squares 482 13.7. Optimization with constraints 483
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> def f(x):
... return np.sqrt((x[0] - 3)**2 + (x[1] - 2)**2) x: [ 1.250e+00 2.500e-01]
>>> sp.optimize.minimize(f, np.array([0, 0]), bounds=((-1.5, 1.5), (-1.5, 1.5))) nit: 5
message: CONVERGENCE: NORM OF PROJECTED GRADIENT <= PGTOL jac: [-7.071e-01 -7.071e-01]
success: True nfev: 15
status: 0 njev: 5
fun: 1.5811388300841898
x: [ 1.500e+00 1.500e+00]
nit: 2 . Warning
jac: [-9.487e-01 -3.162e-01]
nfev: 9 The above problem is known as the Lasso problem in statistics, and there exist very efficient solvers
njev: 3 for it (for instance in scikit-learn). In general do not use generic solvers when specific ones exist.
hess_inv: <2x2 LbfgsInvHessProduct with dtype=float64>

Lagrange multipliers

If you are ready to do a bit of math, many constrained optimization problems can be converted to
non-constrained optimization problems using a mathematical trick known as Lagrange multipliers.

13.8 Full code examples

13.9 Examples for the mathematical optimization chapter

ã See also
13.7.2 General constraints
Other Software
Equality and inequality constraints specified as functions: 𝑓 (𝑥) = 0 and 𝑔(𝑥) < 0.
SciPy tries to include the best well-established, general-use, and permissively-licensed optimization
• scipy.optimize.fmin_slsqp() Sequential least square programming: equality and inequality con- algorithms available. However, even better options for a given task may be available in other libraries;
straints: please also see IPOPT and PyGMO.

>>> def f(x):

... return np.sqrt((x[0] - 3)**2 + (x[1] - 2)**2)

>>> def constraint(x):

... return np.atleast_1d(1.5 - np.sum(np.abs(x)))

>>> x0 = np.array([0, 0])

>>> sp.optimize.minimize(f, x0, constraints={"fun": constraint, "type": "ineq"})
message: Optimization terminated successfully
success: True
status: 0
fun: 2.47487373504...
(continues on next page)

13.7. Optimization with constraints 484 13.8. Full code examples 485
 Scientific Python Lectures, Edition 2025.1rc0.dev0

14.1 Introduction
This chapter covers the following techniques:
• Python-C-Api

14
• Ctypes
• SWIG (Simplified Wrapper and Interface Generator)
• Cython
These four techniques are perhaps the most well known ones, of which Cython is probably the most
advanced one and the one you should consider using first. The others are also important, if you want to
CHAPTER understand the wrapping problem from different angles. Having said that, there are other alternatives
out there, but having understood the basics of the ones above, you will be in a position to evaluate the
technique of your choice to see if it fits your needs.
The following criteria may be useful when evaluating a technology:
• Are additional libraries required?
• Is the code autogenerated?
• Does it need to be compiled?
Interfacing with C • Is there good support for interacting with NumPy arrays?
• Does it support C++?
Before you set out, you should consider your use case. When interfacing with native code, there are
usually two use-cases that come up:
• Existing code in C/C++ that needs to be leveraged, either because it already exists, or because it
is faster.
Author: Valentin Haenel • Python code too slow, push inner loops to native code
This chapter contains an introduction to the many different routes for making your native code (primarily Each technology is demonstrated by wrapping the cos function from math.h. While this is a mostly a
C/C++) available from Python, a process commonly referred to wrapping. The goal of this chapter is to trivial example, it should serve us well to demonstrate the basics of the wrapping solution. Since each
give you a flavour of what technologies exist and what their respective merits and shortcomings are, so technique also includes some form of NumPy support, this is also demonstrated using an example where
that you can select the appropriate one for your specific needs. In any case, once you do start wrapping, the cosine is computed on some kind of array.
you almost certainly will want to consult the respective documentation for your selected technique.
Last but not least, two small warnings:
• All of these techniques may crash (segmentation fault) the Python interpreter, which is (usually)
Chapters contents
due to bugs in the C code.
• Introduction • All the examples have been done on Linux, they should be possible on other operating systems.
• Python-C-Api • You will need a C compiler for most of the examples.
• Ctypes
• SWIG 14.2 Python-C-Api
• Cython The Python-C-API is the backbone of the standard Python interpreter (a.k.a CPython). Using this API
• Summary it is possible to write Python extension module in C and C++. Obviously, these extension modules can,
by virtue of language compatibility, call any function written in C or C++.
• Further Reading and References
When using the Python-C-API, one usually writes much boilerplate code, first to parse the arguments
• Exercises that were given to a function, and later to construct the return type.
Advantages
• Requires no additional libraries
• Lots of low-level control
• Entirely usable from C++
Disadvantages

486 14.1. Introduction 487

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• May require a substantial amount of effort (continued from previous page)

• Much overhead in the code {"cos_func", cos_func, METH_VARARGS, "evaluate the cosine"},
{NULL, NULL, 0, NULL}
• Must be compiled };
• High maintenance cost
# if PY_MAJOR_VERSION >= 3
• No forward compatibility across Python versions as C-API changes /* module initialization */
• Reference count bugs are easy to create and very hard to track down. /* Python version 3*/
static struct PyModuleDef cModPyDem =
{
ò Note PyModuleDef_HEAD_INIT,
"cos_module", "Some documentation",
The Python-C-Api example here serves mainly for didactic reasons. Many of the other techniques -1,
actually depend on this, so it is good to have a high-level understanding of how it works. In 99% of CosMethods
the use-cases you will be better off, using an alternative technique. };

PyMODINIT_FUNC
PyInit_cos_module(void)
ò Note
{
Since reference counting bugs are easy to create and hard to track down, anyone really needing to return PyModule_Create(&cModPyDem);
use the Python C-API should read the section about objects, types and reference counts from the }
official python documentation. Additionally, there is a tool by the name of cpychecker which can help
discover common errors with reference counting. # else

/* module initialization */
14.2.1 Example /* Python version 2 */
PyMODINIT_FUNC
The following C-extension module, make the cos function from the standard math library available to initcos_module(void)
Python: {
/* Example of wrapping cos function from math.h with the Python-C-API. */ (void) Py_InitModule("cos_module", CosMethods);
}
# include <Python.h>
# include <math.h> # endif

/* wrapped cosine function */ As you can see, there is much boilerplate, both to «massage» the arguments and return types into
static PyObject* cos_func(PyObject* self, PyObject* args) place and for the module initialisation. Although some of this is amortised, as the extension grows, the
{ boilerplate required for each function(s) remains.
double value; The standard python build system, setuptools, supports compiling C-extensions via a setup.py file:
double answer;
from setuptools import setup, Extension
/* parse the input, from python float to c double */
if (!PyArg_ParseTuple(args, "d", &value))
return NULL; # define the extension module
/* if the above function returns -1, an appropriate Python exception will cos_module = Extension("cos_module", sources=["cos_module.c"])
* have been set, and the function simply returns NULL
*/ # run the setup
setup(ext_modules=[cos_module])
/* call cos from libm */
answer = cos(value); The setup file is called as follows:

/* construct the output from cos, from c double to python float */ $ cd advanced/interfacing_with_c/python_c_api
return Py_BuildValue("f", answer);
} $ ls
cos_module.c setup.py
/* define functions in module */
static PyMethodDef CosMethods[] = $ python setup.py build_ext --inplace
{ running build_ext
(continues on next page) (continues on next page)

14.2. Python-C-Api 488 14.2. Python-C-Api 489

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) 14.2.2 NumPy Support

building 'cos_module' extension Analog to the Python-C-API, NumPy, which is itself implemented as a C-extension, comes with the
creating build NumPy-C-API. This API can be used to create and manipulate NumPy arrays from C, when writing a
creating build/temp.linux-x86_64-2.7 custom C-extension. See also: Advanced NumPy.
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-
˓→prototypes -fPIC -I/home/esc/anaconda/include/python2.7 -c cos_module.c -o build/

˓→temp.linux-x86_64-2.7/cos_module.o ò Note
gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_module.o -L/home/esc/anaconda/
˓→lib -lpython2.7 -o /home/esc/git-working/scientific-python-lectures/advanced/
If you do ever need to use the NumPy C-API refer to the documentation about Arrays and Iterators.
˓→interfacing_with_c/python_c_api/cos_module.so

The following example shows how to pass NumPy arrays as arguments to functions and how to iterate
$ ls over NumPy arrays using the (old) NumPy-C-API. It simply takes an array as argument applies the
build/ cos_module.c cos_module.so setup.py cosine function from the math.h and returns a resulting new array.

• build_ext is to build extension modules /* Example of wrapping the cos function from math.h using the NumPy-C-API. */
• --inplace will output the compiled extension module into the current directory
# include <Python.h>
The file cos_module.so contains the compiled extension, which we can now load in the IPython inter- # include <numpy/arrayobject.h>
preter: # include <math.h>

ò Note /* wrapped cosine function */

static PyObject* cos_func_np(PyObject* self, PyObject* args)
In Python 3, the filename for compiled modules includes metadata on the Python interpreter (see
{
PEP 3149) and is thus longer. The import statement is not affected by this.
PyArrayObject *arrays[2]; /* holds input and output array */
PyObject *ret;
In [1]: import cos_module NpyIter *iter;
npy_uint32 op_flags[2];
In [2]: cos_module? npy_uint32 iterator_flags;
Type: module PyArray_Descr *op_dtypes[2];
String Form:<module 'cos_module' from 'cos_module.so'>
File: /home/esc/git-working/scientific-python-lectures/advanced/interfacing_ NpyIter_IterNextFunc *iternext;
˓→with_c/python_c_api/cos_module.so

Docstring: <no docstring> /* parse single NumPy array argument */

if (!PyArg_ParseTuple(args, "O!", &PyArray_Type, &arrays[0])) {
In [3]: dir(cos_module) return NULL;
Out[3]: ['__doc__', '__file__', '__name__', '__package__', 'cos_func'] }

In [4]: cos_module.cos_func(1.0) arrays[1] = NULL; /* The result will be allocated by the iterator */
Out[4]: 0.5403023058681398
/* Set up and create the iterator */
In [5]: cos_module.cos_func(0.0) iterator_flags = (NPY_ITER_ZEROSIZE_OK |
Out[5]: 1.0 /*
* Enable buffering in case the input is not behaved
In [6]: cos_module.cos_func(3.14159265359) * (native byte order or not aligned),
Out[6]: -1.0 * disabling may speed up some cases when it is known to
* be unnecessary.
*/
Now let’s see how robust this is:
NPY_ITER_BUFFERED |
In [7]: cos_module.cos_func('foo') /* Manually handle innermost iteration for speed: */
--------------------------------------------------------------------------- NPY_ITER_EXTERNAL_LOOP |
TypeError Traceback (most recent call last) NPY_ITER_GROWINNER);
<ipython-input-10-11bee483665d> in <module>()
----> 1 cos_module.cos_func('foo') op_flags[0] = (NPY_ITER_READONLY |
TypeError: a float is required /*
* Required that the arrays are well behaved, since the cos
* call below requires this.
(continues on next page)

14.2. Python-C-Api 490 14.2. Python-C-Api 491

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

*/ double *out = (double *)dataptr[1];
NPY_ITER_NBO | char *in = dataptr[0];
NPY_ITER_ALIGNED);
/* The output is allocated and guaranteed contiguous (out++ works): */
/* Ask the iterator to allocate an array to write the output to */ assert(strideptr[1] == sizeof(double));
op_flags[1] = NPY_ITER_WRITEONLY | NPY_ITER_ALLOCATE;
/*
/* * For optimization it can make sense to add a check for
* Ensure the iteration has the correct type, could be checked * stride == sizeof(double) to allow the compiler to optimize for that.
* specifically here. */
*/ while (count--) {
op_dtypes[0] = PyArray_DescrFromType(NPY_DOUBLE); *out = cos(*(double *)in);
op_dtypes[1] = op_dtypes[0]; out++;
in += stride;
/* Create the NumPy iterator object: */ }
iter = NpyIter_MultiNew(2, arrays, iterator_flags, } while (iternext(iter));
/* Use input order for output and iteration */
NPY_KEEPORDER, /* Clean up and return the result */
/* Allow only byte-swapping of input */ NpyIter_Deallocate(iter);
NPY_EQUIV_CASTING, op_flags, op_dtypes); return ret;
Py_DECREF(op_dtypes[0]); /* The second one is identical. */ }

if (iter == NULL)
return NULL; /* define functions in module */
static PyMethodDef CosMethods[] =
iternext = NpyIter_GetIterNext(iter, NULL); {
if (iternext == NULL) { {"cos_func_np", cos_func_np, METH_VARARGS,
NpyIter_Deallocate(iter); "evaluate the cosine on a NumPy array"},
return NULL; {NULL, NULL, 0, NULL}
} };

/* Fetch the output array which was allocated by the iterator: */

ret = (PyObject *)NpyIter_GetOperandArray(iter)[1]; # if PY_MAJOR_VERSION >= 3
Py_INCREF(ret); /* module initialization */
/* Python version 3*/
if (NpyIter_GetIterSize(iter) == 0) { static struct PyModuleDef cModPyDem = {
/* PyModuleDef_HEAD_INIT,
* If there are no elements, the loop cannot be iterated. "cos_module", "Some documentation",
* This check is necessary with NPY_ITER_ZEROSIZE_OK. -1,
*/ CosMethods
NpyIter_Deallocate(iter); };
return ret; PyMODINIT_FUNC PyInit_cos_module_np(void) {
} PyObject *module;
module = PyModule_Create(&cModPyDem);
/* The location of the data pointer which the iterator may update */ if(module==NULL) return NULL;
char **dataptr = NpyIter_GetDataPtrArray(iter); /* IMPORTANT: this must be called */
/* The location of the stride which the iterator may update */ import_array();
npy_intp *strideptr = NpyIter_GetInnerStrideArray(iter); if (PyErr_Occurred()) return NULL;
/* The location of the inner loop size which the iterator may update */ return module;
npy_intp *innersizeptr = NpyIter_GetInnerLoopSizePtr(iter); }

/* iterate over the arrays */ # else

do { /* module initialization */
npy_intp stride = strideptr[0]; /* Python version 2 */
npy_intp count = *innersizeptr; PyMODINIT_FUNC initcos_module_np(void) {
/* out is always contiguous, so use double */ PyObject *module;
(continues on next page) (continues on next page)

14.2. Python-C-Api 492 14.2. Python-C-Api 493

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

module = Py_InitModule("cos_module_np", CosMethods);
if(module==NULL) return; # The function fails if the data type is incorrect:
/* IMPORTANT: this must be called */ x_incorrect_dtype = x.astype(np.float32)
import_array(); try:
return; cos_module_np.cos_func_np(x_incorrect_dtype)
} assert 0, "This cannot be reached."
except TypeError:
# endif # A TypeError will be raised, this can be changed by changing the
# casting rule.
To compile this we can use setuptools again. However we need to be sure to include the NumPy headers pass
by using numpy.get_include().
And this should result in the following figure:
from setuptools import setup, Extension
import numpy

# define the extension module

cos_module_np = Extension(
"cos_module_np", sources=["cos_module_np.c"], include_dirs=[numpy.get_include()]
)

# run the setup

setup(ext_modules=[cos_module_np])

To convince ourselves if this does actually works, we run the following test script:

import cos_module_np 14.3 Ctypes

import numpy as np
import matplotlib.pyplot as plt Ctypes is a foreign function library for Python. It provides C compatible data types, and allows calling
functions in DLLs or shared libraries. It can be used to wrap these libraries in pure Python.
x = np.arange(0, 2 * np.pi, 0.1)
y = cos_module_np.cos_func_np(x) Advantages
plt.plot(x, y) • Part of the Python standard library
plt.show()
• Does not need to be compiled
• Wrapping code entirely in Python
# Below are more specific tests for less common usage
# --------------------------------------------------- Disadvantages
• Requires code to be wrapped to be available as a shared library (roughly speaking *.dll in Windows
# The function is OK with `x` not having any elements: *.so in Linux and *.dylib in Mac OSX.)
x_empty = np.array([], dtype=np.float64)
y_empty = cos_module_np.cos_func_np(x_empty) • No good support for C++
assert np.array_equal(y_empty, np.array([], dtype=np.float64))
14.3.1 Example
# The function can handle arbitrary dimensions and non-contiguous data.
As advertised, the wrapper code is in pure Python.
# `x_2d` contains the same values, but has a different shape.
# Note: `x_2d.flags` shows it is not contiguous and `x2.ravel() == x` """Example of wrapping cos function from math.h using ctypes."""
x_2d = x.repeat(2)[::2].reshape(-1, 3)
y_2d = cos_module_np.cos_func_np(x_2d) import ctypes
# When reshaped back, the same result is given:
assert np.array_equal(y_2d.ravel(), y) # find and load the library

# The function handles incorrect byte-order fine: # OSX or linux

x_not_native_byteorder = x.astype(x.dtype.newbyteorder()) from ctypes.util import find_library
y_not_native_byteorder = cos_module_np.cos_func_np(x_not_native_byteorder)
assert np.array_equal(y_not_native_byteorder, y) libm_name = find_library("m")
(continues on next page) (continues on next page)

14.2. Python-C-Api 494 14.3. Ctypes 495

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) As with the previous example, this code is somewhat robust, although the error message is not quite as
assert libm_name is not None, "Cannot find libm (math) on this system :/ That's bad." helpful, since it does not tell us what the type should be.

In [14]: cos_module.cos_func('foo')
libm = ctypes.cdll.LoadLibrary(libm_name)
---------------------------------------------------------------------------
ArgumentError Traceback (most recent call last)
# Windows
<ipython-input-7-11bee483665d> in <module>()
# from ctypes import windll
----> 1 cos_module.cos_func('foo')
# libm = cdll.msvcrt
/home/esc/git-working/scientific-python-lectures/advanced/interfacing_with_c/ctypes/
˓→cos_module.py in cos_func(arg)

12 def cos_func(arg):
# set the argument type
13 ''' Wrapper for cos from math.h '''
libm.cos.argtypes = [ctypes.c_double]
---> 14 return libm.cos(arg)
# set the return type
ArgumentError: argument 1: <type 'exceptions.TypeError'>: wrong type
libm.cos.restype = ctypes.c_double

14.3.2 NumPy Support

def cos_func(arg):
"""Wrapper for cos from math.h""" NumPy contains some support for interfacing with ctypes. In particular there is support for exporting
return libm.cos(arg) certain attributes of a NumPy array as ctypes data-types and there are functions to convert from C
arrays to NumPy arrays and back.
• Finding and loading the library may vary depending on your operating system, check the docu- For more information, consult the corresponding section in the NumPy Cookbook and the API docu-
mentation for details mentation for numpy.ndarray.ctypes and numpy.ctypeslib.
• This may be somewhat deceptive, since the math library exists in compiled form on the system For the following example, let’s consider a C function in a library that takes an input and an output
already. If you were to wrap a in-house library, you would have to compile it first, which may or array, computes the cosine of the input array and stores the result in the output array.
may not require some additional effort.
The library consists of the following header file (although this is not strictly needed for this example, we
We may now use this, as before: list it for completeness):
In [8]: import cos_module void cos_doubles(double * in_array, double * out_array, int size);
In [9]: cos_module?
The function implementation resides in the following C source file:
Type: module
String Form:<module 'cos_module' from 'cos_module.py'> # include <math.h>
File: /home/esc/git-working/scientific-python-lectures/advanced/interfacing_
˓→with_c/ctypes/cos_module.py
/* Compute the cosine of each element in in_array, storing the result in
Docstring: <no docstring> * out_array. */
void cos_doubles(double * in_array, double * out_array, int size){
In [10]: dir(cos_module) int i;
Out[10]: for(i=0;i<size;i++){
['__builtins__', out_array[i] = cos(in_array[i]);
'__doc__', }
'__file__', }
'__name__',
'__package__', And since the library is pure C, we can’t use setuptools to compile it, but must use a combination of
'cos_func', make and gcc:
'ctypes',
'find_library', m.PHONY : clean
'libm']
libcos_doubles.so : cos_doubles.o
In [11]: cos_module.cos_func(1.0) gcc -shared -Wl,-soname,libcos_doubles.so -o libcos_doubles.so cos_doubles.o
Out[11]: 0.5403023058681398
cos_doubles.o : cos_doubles.c
In [12]: cos_module.cos_func(0.0) gcc -c -fPIC cos_doubles.c -o cos_doubles.o
Out[12]: 1.0
clean :
In [13]: cos_module.cos_func(3.14159265359) -rm -vf libcos_doubles.so cos_doubles.o cos_doubles.pyc
Out[13]: -1.0
We can then compile this (on Linux) into the shared library libcos_doubles.so:

14.3. Ctypes 496 14.3. Ctypes 497

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

$ ls
cos_doubles.c cos_doubles.h cos_doubles.py makefile test_cos_doubles.py
$ make
gcc -c -fPIC cos_doubles.c -o cos_doubles.o
gcc -shared -Wl,-soname,libcos_doubles.so -o libcos_doubles.so cos_doubles.o
$ ls
cos_doubles.c cos_doubles.o libcos_doubles.so* test_cos_doubles.py
cos_doubles.h cos_doubles.py makefile

Now we can proceed to wrap this library via ctypes with direct support for (certain kinds of) NumPy
arrays:

"""Example of wrapping a C library function that accepts a C double array as

input using the numpy.ctypeslib.""" 14.4 SWIG
import numpy as np SWIG, the Simplified Wrapper Interface Generator, is a software development tool that connects pro-
import numpy.ctypeslib as npct grams written in C and C++ with a variety of high-level programming languages, including Python.
from ctypes import c_int The important thing with SWIG is, that it can autogenerate the wrapper code for you. While this is an
advantage in terms of development time, it can also be a burden. The generated file tend to be quite
# input type for the cos_doubles function large and may not be too human readable and the multiple levels of indirection which are a result of the
# must be a double array, with single dimension that is contiguous wrapping process, may be a bit tricky to understand.
array_1d_double = npct.ndpointer(dtype=np.double, ndim=1, flags="CONTIGUOUS")
ò Note
# load the library, using NumPy mechanisms
libcd = npct.load_library("libcos_doubles", ".") The autogenerated C code uses the Python-C-Api.

# setup the return types and argument types

libcd.cos_doubles.restype = None Advantages
libcd.cos_doubles.argtypes = [array_1d_double, array_1d_double, c_int] • Can automatically wrap entire libraries given the headers
• Works nicely with C++
def cos_doubles_func(in_array, out_array): Disadvantages
return libcd.cos_doubles(in_array, out_array, len(in_array))
• Autogenerates enormous files
• Note the inherent limitation of contiguous single dimensional NumPy arrays, since the C functions • Hard to debug if something goes wrong
requires this kind of buffer.
• Steep learning curve
• Also note that the output array must be preallocated, for example with numpy.zeros() and the
function will write into it’s buffer.
14.4.1 Example
• Although the original signature of the cos_doubles function is ARRAY, ARRAY, int the final
Let’s imagine that our cos function lives in a cos_module which has been written in c and consists of
cos_doubles_func takes only two NumPy arrays as arguments.
the source file cos_module.c:
And, as before, we convince ourselves that it worked:
# include <math.h>
import numpy as np
import matplotlib.pyplot as plt double cos_func(double arg){
import cos_doubles return cos(arg);
}
x = np.arange(0, 2 * np.pi, 0.1)
y = np.empty_like(x) and the header file cos_module.h:

cos_doubles.cos_doubles_func(x, y) double cos_func(double arg);

plt.plot(x, y)
plt.show() And our goal is to expose the cos_func to Python. To achieve this with SWIG, we must write an
interface file which contains the instructions for SWIG.

/* Example of wrapping cos function from math.h using SWIG. */

(continues on next page)

14.3. Ctypes 498 14.4. SWIG 499

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

In [15]: import cos_module
%module cos_module
%{ In [16]: cos_module?
/* the resulting C file should be built as a python extension */ Type: module
# define SWIG_FILE_WITH_INIT String Form:<module 'cos_module' from 'cos_module.py'>
/* Includes the header in the wrapper code */ File: /home/esc/git-working/scientific-python-lectures/advanced/interfacing_
# include "cos_module.h" ˓→with_c/swig/cos_module.py
%} Docstring: <no docstring>
/* Parse the header file to generate wrappers */
%include "cos_module.h" In [17]: dir(cos_module)
Out[17]:
As you can see, not too much code is needed here. For this simple example it is enough to simply include ['__builtins__',
the header file in the interface file, to expose the function to Python. However, SWIG does allow for '__doc__',
more fine grained inclusion/exclusion of functions found in header files, check the documentation for '__file__',
details. '__name__',
Generating the compiled wrappers is a two stage process: '__package__',
'_cos_module',
1. Run the swig executable on the interface file to generate the files cos_module_wrap.c, which '_newclass',
is the source file for the autogenerated Python C-extension and cos_module.py, which is the '_object',
autogenerated pure python module. '_swig_getattr',
2. Compile the cos_module_wrap.c into the _cos_module.so. Luckily, setuptools knows how to '_swig_property',
handle SWIG interface files, so that our setup.py is simply: '_swig_repr',
'_swig_setattr',
from setuptools import setup, Extension '_swig_setattr_nondynamic',
'cos_func']

setup(ext_modules=[Extension("_cos_module", sources=["cos_module.c", "cos_module.i In [18]: cos_module.cos_func(1.0)

˓→"])]) Out[18]: 0.5403023058681398

$ cd advanced/interfacing_with_c/swig In [19]: cos_module.cos_func(0.0)

Out[19]: 1.0
$ ls
cos_module.c cos_module.h cos_module.i setup.py In [20]: cos_module.cos_func(3.14159265359)
Out[20]: -1.0
$ python setup.py build_ext --inplace
running build_ext Again we test for robustness, and we see that we get a better error message (although, strictly speaking
building '_cos_module' extension in Python there is no double type):
swigging cos_module.i to cos_module_wrap.c
In [21]: cos_module.cos_func('foo')
swig -python -o cos_module_wrap.c cos_module.i
---------------------------------------------------------------------------
creating build
TypeError Traceback (most recent call last)
creating build/temp.linux-x86_64-2.7
<ipython-input-7-11bee483665d> in <module>()
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-
----> 1 cos_module.cos_func('foo')
˓→prototypes -fPIC -I/home/esc/anaconda/include/python2.7 -c cos_module.c -o build/
TypeError: in method 'cos_func', argument 1 of type 'double'
˓→temp.linux-x86_64-2.7/cos_module.o

gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-

˓→prototypes -fPIC -I/home/esc/anaconda/include/python2.7 -c cos_module_wrap.c -o␣
14.4.2 NumPy Support
˓→build/temp.linux-x86_64-2.7/cos_module_wrap.o

gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_module.o build/temp.linux-x86_64- NumPy provides support for SWIG with the numpy.i file. This interface file defines various so-called
˓→2.7/cos_module_wrap.o -L/home/esc/anaconda/lib -lpython2.7 -o /home/esc/git-working/ typemaps which support conversion between NumPy arrays and C-Arrays. In the following example we
˓→scientific-python-lectures/advanced/interfacing_with_c/swig/_cos_module.so will take a quick look at how such typemaps work in practice.
We have the same cos_doubles function as in the ctypes example:
$ ls
build/ cos_module.c cos_module.h cos_module.i cos_module.py _cos_module.so* cos_ void cos_doubles(double * in_array, double * out_array, int size);
˓→module_wrap.c setup.py
# include <math.h>
We can now load and execute the cos_module as we have done in the previous examples:
(continues on next page)

14.4. SWIG 500 14.4. SWIG 501

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

/* Compute the cosine of each element in in_array, storing the result in
* out_array. */
void cos_doubles(double * in_array, double * out_array, int size){ setup(
int i; ext_modules=[
for(i=0;i<size;i++){ Extension(
out_array[i] = cos(in_array[i]); "_cos_doubles",
} sources=["cos_doubles.c", "cos_doubles.i"],
} include_dirs=[numpy.get_include()],
)
This is wrapped as cos_doubles_func using the following SWIG interface file: ]
)
/* Example of wrapping a C function that takes a C double array as input using
* NumPy typemaps for SWIG. */ As previously, we need to use include_dirs to specify the location.

%module cos_doubles $ ls
%{ cos_doubles.c cos_doubles.h cos_doubles.i numpy.i setup.py test_cos_doubles.py
/* the resulting C file should be built as a python extension */ $ python setup.py build_ext -i
# define SWIG_FILE_WITH_INIT running build_ext
/* Includes the header in the wrapper code */ building '_cos_doubles' extension
# include "cos_doubles.h" swigging cos_doubles.i to cos_doubles_wrap.c
%} swig -python -o cos_doubles_wrap.c cos_doubles.i
cos_doubles.i:24: Warning(490): Fragment 'NumPy_Backward_Compatibility' not found.
/* include the NumPy typemaps */ cos_doubles.i:24: Warning(490): Fragment 'NumPy_Backward_Compatibility' not found.
%include "numpy.i" cos_doubles.i:24: Warning(490): Fragment 'NumPy_Backward_Compatibility' not found.
/* need this for correct module initialization */ creating build
%init %{ creating build/temp.linux-x86_64-2.7
import_array(); gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-
%} ˓→prototypes -fPIC -I/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/

˓→include -I/home/esc/anaconda/include/python2.7 -c cos_doubles.c -o build/temp.linux-

/* typemaps for the two arrays, the second will be modified in-place */ ˓→x86_64-2.7/cos_doubles.o

%apply (double* IN_ARRAY1, int DIM1) {(double * in_array, int size_in)} gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-
%apply (double* INPLACE_ARRAY1, int DIM1) {(double * out_array, int size_out)} ˓→prototypes -fPIC -I/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/

˓→include -I/home/esc/anaconda/include/python2.7 -c cos_doubles_wrap.c -o build/temp.

/* Wrapper for cos_doubles that massages the types */ ˓→linux-x86_64-2.7/cos_doubles_wrap.o

%inline %{ In file included from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/

/* takes as input two NumPy arrays */ ˓→include/numpy/ndarraytypes.h:1722,

void cos_doubles_func(double * in_array, int size_in, double * out_array, int␣ from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/
˓→size_out) { ˓→include/numpy/ndarrayobject.h:17,

/* calls the original function, providing only the size of the first */ from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/
cos_doubles(in_array, out_array, size_in); ˓→include/numpy/arrayobject.h:15,

} from cos_doubles_wrap.c:2706:
%} /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/npy_
˓→deprecated_api.h:11:2: warning: #warning "Using deprecated NumPy API, disable it by

• To use the NumPy typemaps, we need include the numpy.i file. ˓→#defining NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION"

gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_doubles.o build/temp.linux-x86_

• Observe the call to import_array() which we encountered already in the NumPy-C-API example. ˓→64-2.7/cos_doubles_wrap.o -L/home/esc/anaconda/lib -lpython2.7 -o /home/esc/git-

• Since the type maps only support the signature ARRAY, SIZE we need to wrap the cos_doubles ˓→working/scientific-python-lectures/advanced/interfacing_with_c/swig_numpy/_cos_

as cos_doubles_func which takes two arrays including sizes as input. ˓→doubles.so

$ ls
• As opposed to the simple SWIG example, we don’t include the cos_doubles.h header, There
build/ cos_doubles.h cos_doubles.py cos_doubles_wrap.c setup.py
is nothing there that we wish to expose to Python since we expose the functionality through cos_doubles.c cos_doubles.i _cos_doubles.so* numpy.i test_cos_doubles.
cos_doubles_func. ˓→py

And, as before we can use setuptools to wrap this:

And, as before, we convince ourselves that it worked:
from setuptools import setup, Extension
import numpy import numpy as np
(continues on next page) (continues on next page)

14.4. SWIG 502 14.4. SWIG 503

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) • Requires an additional library ( but only at build time, at this problem can be overcome by shipping
import matplotlib.pyplot as plt the generated C files)
import cos_doubles
14.5.1 Example
x = np.arange(0, 2 * np.pi, 0.1)
y = np.empty_like(x) The main Cython code for our cos_module is contained in the file cos_module.pyx:

""" Example of wrapping cos function from math.h using Cython. """
cos_doubles.cos_doubles_func(x, y)
plt.plot(x, y) cdef extern from "math.h":
plt.show() double cos(double arg)

def cos_func(arg):
return cos(arg)

Note the additional keywords such as cdef and extern. Also the cos_func is then pure Python.
Again we can use the standard setuptools module, but this time we need some additional pieces from
Cython.Build:

from setuptools import setup, Extension

from Cython.Build import cythonize

extensions = [Extension("cos_module", sources=["cos_module.pyx"])]

14.5 Cython setup(ext_modules=cythonize(extensions))
Cython is both a Python-like language for writing C-extensions and an advanced compiler for this
Compiling this:
language. The Cython language is a superset of Python, which comes with additional constructs that
allow you call C functions and annotate variables and class attributes with c types. In this sense one $ cd advanced/interfacing_with_c/cython
could also call it a Python with types. $ ls
In addition to the basic use case of wrapping native code, Cython supports an additional use-case, cos_module.pyx setup.py
namely interactive optimization. Basically, one starts out with a pure-Python script and incrementally $ python setup.py build_ext --inplace
adds Cython types to the bottleneck code to optimize only those code paths that really matter. running build_ext
cythoning cos_module.pyx to cos_module.c
In this sense it is quite similar to SWIG, since the code can be autogenerated but in a sense it also quite building 'cos_module' extension
similar to ctypes since the wrapping code can (almost) be written in Python. creating build
While others solutions that autogenerate code can be quite difficult to debug (for example SWIG) Cython creating build/temp.linux-x86_64-2.7
comes with an extension to the GNU debugger that helps debug Python, Cython and C code. gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-
˓→prototypes -fPIC -I/home/esc/anaconda/include/python2.7 -c cos_module.c -o build/

˓→temp.linux-x86_64-2.7/cos_module.o

ò Note gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_module.o -L/home/esc/anaconda/

˓→lib -lpython2.7 -o /home/esc/git-working/scientific-python-lectures/advanced/
The autogenerated C code uses the Python-C-Api. ˓→interfacing_with_c/cython/cos_module.so

$ ls
Advantages build/ cos_module.c cos_module.pyx cos_module.so* setup.py

• Python like language for writing C-extensions And running it:

• Autogenerated code
In [22]: import cos_module
• Supports incremental optimization
In [23]: cos_module?
• Includes a GNU debugger extension
Type: module
• Support for C++ (Since version 0.13) String Form:<module 'cos_module' from 'cos_module.so'>
File: /home/esc/git-working/scientific-python-lectures/advanced/interfacing_
Disadvantages
˓→with_c/cython/cos_module.so

• Must be compiled Docstring: <no docstring>

(continues on next page)

14.5. Cython 504 14.5. Cython 505

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

In [24]: dir(cos_module) /* Compute the cosine of each element in in_array, storing the result in
Out[24]: * out_array. */
['__builtins__', void cos_doubles(double * in_array, double * out_array, int size){
'__doc__', int i;
'__file__', for(i=0;i<size;i++){
'__name__', out_array[i] = cos(in_array[i]);
'__package__', }
'__test__', }
'cos_func']
This is wrapped as cos_doubles_func using the following Cython code:
In [25]: cos_module.cos_func(1.0)
Out[25]: 0.5403023058681398 """ Example of wrapping a C function that takes C double arrays as input using
the NumPy declarations from Cython """
In [26]: cos_module.cos_func(0.0)
Out[26]: 1.0 # cimport the Cython declarations for NumPy
cimport numpy as np
In [27]: cos_module.cos_func(3.14159265359)
Out[27]: -1.0 # if you want to use the NumPy-C-API from Cython
# (not strictly necessary for this example, but good practice)
And, testing a little for robustness, we can see that we get good error messages: np.import_array()

In [28]: cos_module.cos_func('foo') # cdefine the signature of our c function

--------------------------------------------------------------------------- cdef extern from "cos_doubles.h":
TypeError Traceback (most recent call last) void cos_doubles (double * in_array, double * out_array, int size)
<ipython-input-7-11bee483665d> in <module>()
----> 1 cos_module.cos_func('foo') # create the wrapper code, with NumPy type annotations
/home/esc/git-working/scientific-python-lectures/advanced/interfacing_with_c/cython/ def cos_doubles_func(np.ndarray[double, ndim=1, mode="c"] in_array not None,
˓→cos_module.so in cos_module.cos_func (cos_module.c:506)() np.ndarray[double, ndim=1, mode="c"] out_array not None):
TypeError: a float is required cos_doubles(<double*> np.PyArray_DATA(in_array),
<double*> np.PyArray_DATA(out_array),
Additionally, it is worth noting that Cython ships with complete declarations for the C math library, in_array.shape[0])
which simplifies the code above to become:
And can be compiled using setuptools:
""" Simpler example of wrapping cos function from math.h using Cython. """
from setuptools import setup, Extension
from libc.math cimport cos from Cython.Build import cythonize
import numpy
def cos_func(arg):
return cos(arg)
extensions = [
In this case the cimport statement is used to import the cos function. Extension(
"cos_doubles",
sources=["_cos_doubles.pyx", "cos_doubles.c"],
14.5.2 NumPy Support include_dirs=[numpy.get_include()],
Cython has support for NumPy via the numpy.pyx file which allows you to add the NumPy array type )
to your Cython code. I.e. like specifying that variable i is of type int, you can specify that variable a ]
is of type numpy.ndarray with a given dtype. Also, certain optimizations such as bounds checking are
supported. Look at the corresponding section in the Cython documentation. In case you want to pass setup(ext_modules=cythonize(extensions))
NumPy arrays as C arrays to your Cython wrapped C functions, there is a section about this in the
Cython documentation. • As with the previous compiled NumPy examples, we need the include_dirs option.
In the following example, we will show how to wrap the familiar cos_doubles function using Cython. $ ls
cos_doubles.c cos_doubles.h _cos_doubles.pyx setup.py test_cos_doubles.py
void cos_doubles(double * in_array, double * out_array, int size);
$ python setup.py build_ext -i
running build_ext
# include <math.h> cythoning _cos_doubles.pyx to _cos_doubles.c
building 'cos_doubles' extension
(continues on next page) (continues on next page)

14.5. Cython 506 14.5. Cython 507

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) 14.6 Summary

creating build
creating build/temp.linux-x86_64-2.7 In this section four different techniques for interfacing with native code have been presented. The table
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict- below roughly summarizes some of the aspects of the techniques.
˓→prototypes -fPIC -I/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/

˓→include -I/home/esc/anaconda/include/python2.7 -c _cos_doubles.c -o build/temp.

x Part of CPython Compiled Autogenerated NumPy Support
˓→linux-x86_64-2.7/_cos_doubles.o

In file included from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/ Python-C-API True True False True

˓→include/numpy/ndarraytypes.h:1722, Ctypes True False False True
from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/ Swig False True True True
˓→include/numpy/ndarrayobject.h:17, Cython False True True True
from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/
˓→include/numpy/arrayobject.h:15,
Of all three presented techniques, Cython is the most modern and advanced. In particular, the ability
from _cos_doubles.c:253: to optimize code incrementally by adding types to your Python code is unique.
/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/npy_
˓→deprecated_api.h:11:2: warning: #warning "Using deprecated NumPy API, disable it by

˓→#defining NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION" 14.7 Further Reading and References

/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/__ufunc_api.
˓→h:236: warning: ‘_import_umath’ defined but not used • Gaël Varoquaux’s blog post about avoiding data copies provides some insight on how to handle
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict- memory management cleverly. If you ever run into issues with large datasets, this is a reference to
˓→prototypes -fPIC -I/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/ come back to for some inspiration.
˓→include -I/home/esc/anaconda/include/python2.7 -c cos_doubles.c -o build/temp.linux-

˓→x86_64-2.7/cos_doubles.o

gcc -pthread -shared build/temp.linux-x86_64-2.7/_cos_doubles.o build/temp.linux-x86_ 14.8 Exercises

˓→64-2.7/cos_doubles.o -L/home/esc/anaconda/lib -lpython2.7 -o /home/esc/git-working/

˓→scientific-python-lectures/advanced/interfacing_with_c/cython_numpy/cos_doubles.so
Since this is a brand new section, the exercises are considered more as pointers as to what to look at
$ ls next, so pick the ones that you find more interesting. If you have good ideas for exercises, please let us
build/ _cos_doubles.c cos_doubles.c cos_doubles.h _cos_doubles.pyx cos_doubles. know!
˓→so* setup.py test_cos_doubles.py 1. Download the source code for each example and compile and run them on your machine.

And, as before, we convince ourselves that it worked: 2. Make trivial changes to each example and convince yourself that this works. ( E.g. change cos for
sin.)
import numpy as np
3. Most of the examples, especially the ones involving NumPy may still be fragile and respond badly
import matplotlib.pyplot as plt
to input errors. Look for ways to crash the examples, figure what the problem is and devise a
import cos_doubles
potential solution. Here are some ideas:
x = np.arange(0, 2 * np.pi, 0.1) 1. Numerical overflow.
y = np.empty_like(x)
2. Input and output arrays that have different lengths.
cos_doubles.cos_doubles_func(x, y) 3. Multidimensional array.
plt.plot(x, y)
4. Empty array
plt.show()
5. Arrays with non-double types
4. Use the %timeit IPython magic to measure the execution time of the various solutions

14.8.1 Python-C-API
1. Modify the NumPy example such that the function takes two input arguments, where the second
is the preallocated output array, making it similar to the other NumPy examples.
2. Modify the example such that the function only takes a single input array and modifies this in
place.
3. Try to fix the example to use the new NumPy iterator protocol. If you manage to obtain a working
solution, please submit a pull-request on github.
4. You may have noticed, that the NumPy-C-API example is the only NumPy example that does not
wrap cos_doubles but instead applies the cos function directly to the elements of the NumPy
array. Does this have any advantages over the other techniques.

14.5. Cython 508 14.6. Summary 509

 Scientific Python Lectures, Edition 2025.1rc0.dev0

5. Can you wrap cos_doubles using only the NumPy-C-API. You may need to ensure that the arrays
have the correct type, are one dimensional and contiguous in memory.

14.8.2 Ctypes
1. Modify the NumPy example such that cos_doubles_func handles the preallocation for you, thus
making it more like the NumPy-C-API example.

14.8.3 SWIG
1. Look at the code that SWIG autogenerates, how much of it do you understand?
2. Modify the NumPy example such that cos_doubles_func handles the preallocation for you, thus
making it more like the NumPy-C-API example.
3. Modify the cos_doubles C function so that it returns an allocated array. Can you wrap this using
SWIG typemaps? If not, why not? Is there a workaround for this specific situation? (Hint: you
know the size of the output array, so it may be possible to construct a NumPy array from the
returned double *.)

14.8.4 Cython Part III

1. Look at the code that Cython autogenerates. Take a closer look at some of the comments that
Cython inserts. What do you see?
2. Look at the section Working with NumPy from the Cython documentation to learn how to incre-
mentally optimize a pure python script that uses NumPy. Packages and applications
3. Modify the NumPy example such that cos_doubles_func handles the preallocation for you, thus
making it more like the NumPy-C-API example.

14.8. Exercises 510 511

 Scientific Python Lectures, Edition 2025.1rc0.dev0

This part of the Scientific Python Lectures is dedicated to various scientific packages useful for extended
needs.

CHAPTER 15
Statistics in Python

Author: Gaël Varoquaux

Requirements

• Standard scientific Python environment (NumPy, SciPy, matplotlib)

• Pandas
• Statsmodels
• Seaborn
To install Python and these dependencies, we recommend that you download Anaconda Python or,
preferably, use the package manager if you are under Ubuntu or other linux.

ã See also

• Bayesian statistics in Python: This chapter does not cover tools for Bayesian statistics. Of
particular interest for Bayesian modelling is PyMC, which implements a probabilistic program-
ming language in Python.
• Read a statistics book: The Think stats book is available as free PDF or in print and is a
great introduction to statistics.

512 513
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

columns giving the different attributes of the data, and rows the observations. For instance, the data
 Tip contained in examples/brain_size.csv:
Why Python for statistics? "";"Gender";"FSIQ";"VIQ";"PIQ";"Weight";"Height";"MRI_Count"
R is a language dedicated to statistics. Python is a general-purpose language with statistics modules. "1";"Female";133;132;124;"118";"64.5";816932
R has more statistical analysis features than Python, and specialized syntaxes. However, when it "2";"Male";140;150;124;".";"72.5";1001121
comes to building complex analysis pipelines that mix statistics with e.g. image analysis, text mining, "3";"Male";139;123;150;"143";"73.3";1038437
or control of a physical experiment, the richness of Python is an invaluable asset. "4";"Male";133;129;128;"172";"68.8";965353
"5";"Female";137;132;134;"147";"65.0";951545

Contents 15.1.2 The pandas data-frame

• Data representation and interaction
 Tip
– Data as a table
– The pandas data-frame We will store and manipulate this data in a pandas.DataFrame, from the pandas module. It is the
Python equivalent of the spreadsheet table. It is different from a 2D numpy array as it has named
• Hypothesis testing: comparing two groups columns, can contain a mixture of different data types by column, and has elaborate selection and
– Student’s t-test: the simplest statistical test pivotal mechanisms.

– Paired tests: repeated measurements on the same individuals

Creating dataframes: reading data files or converting arrays
• Linear models, multiple factors, and analysis of variance
– “formulas” to specify statistical models in Python
Separator
– Multiple Regression: including multiple factors
It is a CSV file, but the separator is “;”
– Post-hoc hypothesis testing: analysis of variance (ANOVA)
• More visualization: seaborn for statistical exploration Reading from a CSV file: Using the above CSV file that gives observations of brain size and weight
and IQ (Willerman et al. 1991), the data are a mixture of numerical and categorical values:
– Pairplot: scatter matrices
– lmplot: plotting a univariate regression >>> import pandas
>>> data = pandas.read_csv('examples/brain_size.csv', sep=';', na_values=".")
• Testing for interactions >>> data
• Full code for the figures Unnamed: 0 Gender FSIQ VIQ PIQ Weight Height MRI_Count
0 1 Female 133 132 124 118.0 64.5 816932
• Solutions to this chapter’s exercises 1 2 Male 140 150 124 NaN 72.5 1001121
2 3 Male 139 123 150 143.0 73.3 1038437
3 4 Male 133 129 128 172.0 68.8 965353
4 5 Female 137 132 134 147.0 65.0 951545
 Tip ...
In this document, the Python inputs are represented with the sign “>>>”.

. Warning

Missing values
The weight of the second individual is missing in the CSV file. If we don’t specify the missing value
Disclaimer: Gender questions (NA = not available) marker, we will not be able to do statistical analysis.
Some of the examples of this tutorial are chosen around gender questions. The reason is that on such
questions controlling the truth of a claim actually matters to many people.

15.1 Data representation and interaction

Creating from arrays: A pandas.DataFrame can also be seen as a dictionary of 1D ‘series’, eg arrays
or lists. If we have 3 numpy arrays:
15.1.1 Data as a table
The setting that we consider for statistical analysis is that of multiple observations or samples described
by a set of different attributes or features. The data can than be seen as a 2D table, or matrix, with

15.1. Data representation and interaction 514 15.1. Data representation and interaction 515
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

>>> import numpy as np ò Note

>>> t = np.linspace(-6, 6, 20)
>>> sin_t = np.sin(t) For a quick view on a large dataframe, use its describe method: pandas.DataFrame.describe().
>>> cos_t = np.cos(t)

We can expose them as a pandas.DataFrame:

>>> pandas.DataFrame({'t': t, 'sin': sin_t, 'cos': cos_t})

t sin cos
0 -6.000000 0.279415 0.960170 groupby: splitting a dataframe on values of categorical variables:
1 -5.368421 0.792419 0.609977
>>> groupby_gender = data.groupby('Gender')
2 -4.736842 0.999701 0.024451
>>> for gender, value in groupby_gender['VIQ']:
3 -4.105263 0.821291 -0.570509
... print((gender, value.mean()))
4 -3.473684 0.326021 -0.945363
('Female', np.float64(109.45))
5 -2.842105 -0.295030 -0.955488
('Male', np.float64(115.25))
6 -2.210526 -0.802257 -0.596979
7 -1.578947 -0.999967 -0.008151
8 -0.947368 -0.811882 0.583822 groupby_gender is a powerful object that exposes many operations on the resulting group of dataframes:
... >>> groupby_gender.mean()
Unnamed: 0 FSIQ VIQ PIQ Weight Height MRI_Count
Gender
Female 19.65 111.9 109.45 110.45 137.200000 65.765000 862654.6
Male 21.35 115.0 115.25 111.60 166.444444 71.431579 954855.4

Other inputs: pandas can input data from SQL, excel files, or other formats. See the pandas docu-
mentation.  Tip

Use tab-completion on groupby_gender to find more. Other common grouping functions are median,
count (useful for checking to see the amount of missing values in different subsets) or sum. Groupby
evaluation is lazy, no work is done until an aggregation function is applied.

Manipulating data
data is a pandas.DataFrame, that resembles R’s dataframe:

>>> data.shape # 40 rows and 8 columns

(40, 8)

>>> data.columns # It has columns

Index(['Unnamed: 0', 'Gender', 'FSIQ', 'VIQ', 'PIQ', 'Weight', 'Height',
'MRI_Count'],
dtype='object')

>>> print(data['Gender']) # Columns can be addressed by name

0 Female
1 Male
2 Male
3 Male
4 Female
...
Exercise
>>> # Simpler selector
>>> data[data['Gender'] == 'Female']['VIQ'].mean() • What is the mean value for VIQ for the full population?
np.float64(109.45) • How many males/females were included in this study?
Hint use ‘tab completion’ to find out the methods that can be called, instead of ‘mean’ in the
above example.

15.1. Data representation and interaction 516 15.1. Data representation and interaction 517
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• What is the average value of MRI counts expressed in log units, for males and females? (continued from previous page)
<Axes: xlabel='FSIQ', ylabel='VIQ'>],
[<Axes: xlabel='PIQ', ylabel='FSIQ'>,
<Axes: xlabel='VIQ', ylabel='FSIQ'>,
ò Note
<Axes: xlabel='FSIQ', ylabel='FSIQ'>]], dtype=object)
groupby_gender.boxplot is used for the plots above (see this example).

Two populations

The IQ metrics are bimodal, as if there are 2 sub-populations.

Plotting data
Pandas comes with some plotting tools (pandas.plotting, using matplotlib behind the scene) to display
statistics of the data in dataframes:
Scatter matrices:

>>> from pandas import plotting

>>> plotting.scatter_matrix(data[['Weight', 'Height', 'MRI_Count']])
array([[<Axes: xlabel='Weight', ylabel='Weight'>,
<Axes: xlabel='Height', ylabel='Weight'>,
<Axes: xlabel='MRI_Count', ylabel='Weight'>],
[<Axes: xlabel='Weight', ylabel='Height'>,
<Axes: xlabel='Height', ylabel='Height'>,
<Axes: xlabel='MRI_Count', ylabel='Height'>],
[<Axes: xlabel='Weight', ylabel='MRI_Count'>,
<Axes: xlabel='Height', ylabel='MRI_Count'>,
<Axes: xlabel='MRI_Count', ylabel='MRI_Count'>]], dtype=object)

Exercise

Plot the scatter matrix for males only, and for females only. Do you think that the 2 sub-populations
correspond to gender?

15.2 Hypothesis testing: comparing two groups

For simple statistical tests, we will use the scipy.stats sub-module of SciPy:

>>> import scipy as sp

ã See also

SciPy is a vast library. For a quick summary to the whole library, see the scipy chapter.

>>> plotting.scatter_matrix(data[['PIQ', 'VIQ', 'FSIQ']]) 15.2.1 Student’s t-test: the simplest statistical test
array([[<Axes: xlabel='PIQ', ylabel='PIQ'>,
<Axes: xlabel='VIQ', ylabel='PIQ'>,
<Axes: xlabel='FSIQ', ylabel='PIQ'>],
[<Axes: xlabel='PIQ', ylabel='VIQ'>,
<Axes: xlabel='VIQ', ylabel='VIQ'>,
(continues on next page)

15.1. Data representation and interaction 518 15.2. Hypothesis testing: comparing two groups 519
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

One-sample tests: testing the value of a population mean

>>> sp.stats.mannwhitneyu(female_viq, male_viq)
MannwhitneyuResult(statistic=np.float64(164.5), pvalue=np.float64(0.34228868687...))

15.2.2 Paired tests: repeated measurements on the same individuals

PIQ, VIQ, and FSIQ give three measures of IQ. Let us test whether FISQ and PIQ are significantly
different. We can use an “independent sample” test:
scipy.stats.ttest_1samp() tests the null hypothesis that the mean of the population underlying the
data is equal to a given value. It returns the T statistic, and the p-value (see the function’s help): >>> sp.stats.ttest_ind(data['FSIQ'], data['PIQ'])
TtestResult(statistic=np.float64(0.46563759638...), pvalue=np.float64(0.64277250...),␣
>>> sp.stats.ttest_1samp(data['VIQ'], 0) ˓→df=np.float64(78.0))
TtestResult(statistic=np.float64(30.088099970...), pvalue=np.float64(1.32891964...e-
˓→28), df=np.int64(39))
The problem with this approach is that it ignores an important relationship between observations: FSIQ
and PIQ are measured on the same individuals. Thus, the variance due to inter-subject variability is
The p-value of 10− 28 indicates that such an extreme value of the statistic is unlikely to be observed confounding, reducing the power of the test. This variability can be removed using a “paired test” or
under the null hypothesis. This may be taken as evidence that the null hypothesis is false and that the “repeated measures test”:
population mean IQ (VIQ measure) is not 0.
Technically, the p-value of the t-test is derived under the assumption that the means of samples drawn >>> sp.stats.ttest_rel(data['FSIQ'], data['PIQ'])
from the population are normally distributed. This condition is exactly satisfied when the population TtestResult(statistic=np.float64(1.784201940...), pvalue=np.float64(0.082172638183...
˓→), df=np.int64(39))
itself is normally distributed; however, due to the central limit theorem, the condition is nearly true for
reasonably large samples drawn from populations that follow a variety of non-normal distributions.
Nonetheless, if we are concerned that violation of the normality assumptions will affect the conclusions
of the test, we can use a Wilcoxon signed-rank test, which relaxes this assumption at the expense of test
power:

>>> sp.stats.wilcoxon(data['VIQ'])
WilcoxonResult(statistic=np.float64(0.0), pvalue=np.float64(3.4881726...e-08))

Two-sample t-test: testing for difference across populations

We have seen above that the mean VIQ in the male and female samples were different. To test whether
this difference is significant (and suggests that there is a difference in population means), we perform a
two-sample t-test using scipy.stats.ttest_ind(): This is equivalent to a one-sample test on the differences between paired observations:

>>> female_viq = data[data['Gender'] == 'Female']['VIQ'] >>> sp.stats.ttest_1samp(data['FSIQ'] - data['PIQ'], 0)

>>> male_viq = data[data['Gender'] == 'Male']['VIQ'] TtestResult(statistic=np.float64(1.784201940...), pvalue=np.float64(0.082172638...),␣
˓→df=np.int64(39))
>>> sp.stats.ttest_ind(female_viq, male_viq)
TtestResult(statistic=np.float64(-0.77261617232...), pvalue=np.float64(0.
˓→4445287677858...), df=np.float64(38.0))
Accordingly, we can perform a nonparametric version of the test with wilcoxon.

>>> sp.stats.wilcoxon(data['FSIQ'], data['PIQ'], method="approx")

The corresponding non-parametric test is the Mann–Whitney U test, scipy.stats.mannwhitneyu(). WilcoxonResult(statistic=np.float64(274.5), pvalue=np.float64(0.106594927135...))

15.2. Hypothesis testing: comparing two groups 520 15.2. Hypothesis testing: comparing two groups 521
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Exercise “formulas” for statistics in Python

• Test the difference between weights in males and females. See the statsmodels documentation
• Use non parametric statistics to test the difference between VIQ in males and females.
Conclusion: we find that the data does not support the hypothesis that males and females have
different VIQ.

Then we specify an OLS model and fit it:

>>> from statsmodels.formula.api import ols

>>> model = ols("y ~ x", data).fit()

15.3 Linear models, multiple factors, and analysis of variance We can inspect the various statistics derived from the fit:

15.3.1 “formulas” to specify statistical models in Python >>> print(model.summary())

OLS Regression Results
A simple linear regression ==============================================================================
Dep. Variable: y R-squared: 0.901
Model: OLS Adj. R-squared: 0.896
Method: Least Squares F-statistic: 164.5
Date: ... Prob (F-statistic): 1.72e-10
Time: ... Log-Likelihood: -51.758
No. Observations: 20 AIC: 107.5
Df Residuals: 18 BIC: 109.5
Df Model: 1
Covariance Type: nonrobust
==============================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
Intercept -4.2948 0.759 -5.661 0.000 -5.889 -2.701
x 3.2060 0.250 12.825 0.000 2.681 3.731
==============================================================================
Omnibus: 1.218 Durbin-Watson: 1.796
Given two set of observations, x and y, we want to test the hypothesis that y is a linear function of x. Prob(Omnibus): 0.544 Jarque-Bera (JB): 0.999
In other terms: Skew: 0.503 Prob(JB): 0.607
𝑦 = 𝑥 * coef + intercept + 𝑒 Kurtosis: 2.568 Cond. No. 3.03
==============================================================================
where e is observation noise. We will use the statsmodels module to:
1. Fit a linear model. We will use the simplest strategy, ordinary least squares (OLS). Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly␣
2. Test that coef is non zero. ˓→specified.

Terminology:

Statsmodels uses a statistical terminology: the y variable in statsmodels is called ‘endogenous’ while
First, we generate simulated data according to the model: the x variable is called exogenous. This is discussed in more detail here.
>>> import numpy as np To simplify, y (endogenous) is the value you are trying to predict, while x (exogenous) represents the
>>> x = np.linspace(-5, 5, 20) features you are using to make the prediction.
>>> rng = np.random.default_rng(27446968)
>>> # normal distributed noise
>>> y = -5 + 3*x + 4 * rng.normal(size=x.shape) Exercise
>>> # Create a data frame containing all the relevant variables
>>> data = pandas.DataFrame({'x': x, 'y': y}) Retrieve the estimated parameters from the model above. Hint: use tab-completion to find the
relevant attribute.

15.3. Linear models, multiple factors, and analysis of variance 522 15.3. Linear models, multiple factors, and analysis of variance 523
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Categorical variables: comparing groups or multiple categories

Link to t-tests between different FSIQ and PIQ
Let us go back the data on brain size:
To compare different types of IQ, we need to create a “long-form” table, listing IQs, where the type
>>> data = pandas.read_csv('examples/brain_size.csv', sep=';', na_values=".") of IQ is indicated by a categorical variable:
>>>data_fisq = pandas.DataFrame({'iq': data['FSIQ'], 'type': 'fsiq'})
We can write a comparison between IQ of male and female using a linear model:
>>>data_piq = pandas.DataFrame({'iq': data['PIQ'], 'type': 'piq'})
>>> model = ols("VIQ ~ Gender + 1", data).fit() >>>data_long = pandas.concat((data_fisq, data_piq))
>>> print(model.summary()) >>>print(data_long)
OLS Regression Results iq type
============================================================================== 0 133 fsiq
Dep. Variable: VIQ R-squared: 0.015 1 140 fsiq
Model: OLS Adj. R-squared: -0.010 2 139 fsiq
Method: Least Squares F-statistic: 0.5969 3 133 fsiq
Date: ... Prob (F-statistic): 0.445 4 137 fsiq
Time: ... Log-Likelihood: -182.42 ... ... ...
No. Observations: 40 AIC: 368.8 35 128 piq
Df Residuals: 38 BIC: 372.2 36 124 piq
Df Model: 1 37 94 piq
Covariance Type: nonrobust 38 74 piq
================================================================================== 39 89 piq
coef std err t P>|t| [0.025 0.975]
---------------------------------------------------------------------------------- [80 rows x 2 columns]
Intercept 109.4500 5.308 20.619 0.000 98.704 120.196
Gender[T.Male] 5.8000 7.507 0.773 0.445 -9.397 20.997 >>> model = ols("iq ~ type", data_long).fit()
============================================================================== >>> print(model.summary())
Omnibus: 26.188 Durbin-Watson: 1.709 OLS Regression Results
Prob(Omnibus): 0.000 Jarque-Bera (JB): 3.703 ...
Skew: 0.010 Prob(JB): 0.157 ==========================...
Kurtosis: 1.510 Cond. No. 2.62 coef std err t P>|t| [0.025 0.975]
============================================================================== ------------------------------------------...
Intercept 113.4500 3.683 30.807 0.000 106.119 120.781
Notes: type[T.piq] -2.4250 5.208 -0.466 0.643 -12.793 7.943
[1] Standard Errors assume that the covariance matrix of the errors is correctly␣ ...
˓→specified.
We can see that we retrieve the same values for t-test and corresponding p-values for the effect of the
type of iq than the previous t-test:
Tips on specifying model >>> sp.stats.ttest_ind(data['FSIQ'], data['PIQ'])
TtestResult(statistic=np.float64(0.46563759638...), pvalue=np.float64(0.64277250...
Forcing categorical: the ‘Gender’ is automatically detected as a categorical variable, and thus each ˓→), df=np.float64(78.0))
of its different values are treated as different entities.
An integer column can be forced to be treated as categorical using:
>>> model = ols('VIQ ~ C(Gender)', data).fit()

Intercept: We can remove the intercept using - 1 in the formula, or force the use of an intercept
using + 1.

 Tip

By default, statsmodels treats a categorical variable with K possible values as K-1 ‘dummy’
boolean variables (the last level being absorbed into the intercept term). This is almost always a
good default choice - however, it is possible to specify different encodings for categorical variables
(https://www.statsmodels.org/devel/contrasts.html).

15.3. Linear models, multiple factors, and analysis of variance 524 15.3. Linear models, multiple factors, and analysis of variance 525
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

15.3.2 Multiple Regression: including multiple factors

Consider a linear model explaining a variable z (the dependent variable) with 2 variables x and y:
𝑧 = 𝑥 𝑐1 + 𝑦 𝑐2 + 𝑖 + 𝑒
Such a model can be seen in 3D as fitting a plane to a cloud of (x, y, z) points.
>>> data = pandas.read_csv('examples/iris.csv')
>>> model = ols('sepal_width ~ name + petal_length', data).fit()
>>> print(model.summary())
OLS Regression Results
==========================...
Dep. Variable: sepal_width R-squared: 0.478
Example: the iris data (examples/iris.csv) Model: OLS Adj. R-squared: 0.468
Method: Least Squares F-statistic: 44.63
Date: ... Prob (F-statistic): 1.58e-20
 Tip Time: ... Log-Likelihood: -38.185
No. Observations: 150 AIC: 84.37
Sepal and petal size tend to be related: bigger flowers are bigger! But is there in addition a systematic
Df Residuals: 146 BIC: 96.41
effect of species?
Df Model: 3
Covariance Type: nonrobust
==========================...
coef std err t P>|t| [0.025 0.975]
------------------------------------------...
Intercept 2.9813 0.099 29.989 0.000 2.785 3.178
name[T.versicolor] -1.4821 0.181 -8.190 0.000 -1.840 -1.124
name[T.virginica] -1.6635 0.256 -6.502 0.000 -2.169 -1.158
petal_length 0.2983 0.061 4.920 0.000 0.178 0.418
==========================...
Omnibus: 2.868 Durbin-Watson: 1.753
Prob(Omnibus): 0.238 Jarque-Bera (JB): 2.885
Skew: -0.082 Prob(JB): 0.236
Kurtosis: 3.659 Cond. No. 54.0
==========================...

Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly␣
˓→specified.

15.3. Linear models, multiple factors, and analysis of variance 526 15.3. Linear models, multiple factors, and analysis of variance 527
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

15.3.3 Post-hoc hypothesis testing: analysis of variance (ANOVA)

In the above iris example, we wish to test if the petal length is different between versicolor and virginica,
after removing the effect of sepal width. This can be formulated as testing the difference between the
coefficient associated to versicolor and virginica in the linear model estimated above (it is an Analysis of
Variance, ANOVA). For this, we write a vector of ‘contrast’ on the parameters estimated: we want
to test "name[T.versicolor] - name[T.virginica]", with an F-test:

>>> print(model.f_test([0, 1, -1, 0]))

<F test: F=3.24533535..., p=0.07369..., df_denom=146, df_num=1>

Is this difference significant?

Exercise

Going back to the brain size + IQ data, test if the VIQ of male and female are different after removing
the effect of brain size, height and weight.

15.4 More visualization: seaborn for statistical exploration

Categorical variables can be plotted as the hue:
Seaborn combines simple statistical fits with plotting on pandas dataframes.
>>> seaborn.pairplot(data, vars=['WAGE', 'AGE', 'EDUCATION'],
Let us consider a data giving wages and many other personal information on 500 individuals (Berndt, ... kind='reg', hue='SEX')
ER. The Practice of Econometrics. 1991. NY: Addison-Wesley).

 Tip

The full code loading and plotting of the wages data is found in corresponding example.

>>> print(data)
EDUCATION SOUTH SEX EXPERIENCE UNION WAGE AGE RACE \
0 8 0 1 21 0 0.707570 35 2
1 9 0 1 42 0 0.694605 57 3
2 12 0 0 1 0 0.824126 19 3
3 12 0 0 4 0 0.602060 22 3
...

15.4.1 Pairplot: scatter matrices

We can easily have an intuition on the interactions between continuous variables using seaborn.
pairplot() to display a scatter matrix:

>>> import seaborn

>>> seaborn.pairplot(data, vars=['WAGE', 'AGE', 'EDUCATION'],
... kind='reg')

15.4. More visualization: seaborn for statistical exploration 528 15.4. More visualization: seaborn for statistical exploration 529
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

15.4.2 lmplot: plotting a univariate regression

A regression capturing the relation between one variable and another, eg wage, and education, can be
plotted using seaborn.lmplot():

>>> seaborn.lmplot(y='WAGE', x='EDUCATION', data=data)

Robust regression

Look and feel and matplotlib settings

 Tip
Seaborn changes the default of matplotlib figures to achieve a more “modern”, “excel-like” look. It
does that upon import. You can reset the default using: Given that, in the above plot, there seems to be a couple of data points that are outside of the
main cloud to the right, they might be outliers, not representative of the population, but driving
>>> import matplotlib.pyplot as plt
the regression.
>>> plt.rcdefaults()

To compute a regression that is less sensitive to outliers, one must use a robust model. This is done
in seaborn using robust=True in the plotting functions, or in statsmodels by replacing the use of the
 Tip OLS by a “Robust Linear Model”, statsmodels.formula.api.rlm().
To switch back to seaborn settings, or understand better styling in seaborn, see the relevant
section of the seaborn documentation.

15.4. More visualization: seaborn for statistical exploration 530 15.4. More visualization: seaborn for statistical exploration 531
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

15.5 Testing for interactions • Conditionning (adding factors that can explain all or part of the variation) is an important
modeling aspect that changes the interpretation.

15.6 Full code for the figures

Code examples for the statistics chapter.

15.6.1 Boxplots and paired differences

Plot boxplots for FSIQ, PIQ, and the paired difference between the two: while the spread (error bars)
for FSIQ and PIQ are very large, there is a systematic (common) effect due to the subjects. This effect
is cancelled out in the difference and the spread of the difference (“paired” by subject) is much smaller
than the spread of the individual measures.

Do wages increase more with education for males than females?

 Tip

The plot above is made of two different fits. We need to formulate a single model that tests for a
variance of slope across the two populations. This is done via an “interaction”.

>>> result = sm.ols(formula='wage ~ education + gender + education * gender',

... data=data).fit()
>>> print(result.summary())
...
coef std err t P>|t| [0.025 0.975] •
------------------------------------------------------------------------------
Intercept 0.2998 0.072 4.173 0.000 0.159 0.441
gender[T.male] 0.2750 0.093 2.972 0.003 0.093 0.457
education 0.0415 0.005 7.647 0.000 0.031 0.052
education:gender[T.male] -0.0134 0.007 -1.919 0.056 -0.027 0.000
==========================...
...

Can we conclude that education benefits males more than females?

Take home messages

• Hypothesis testing and p-values give you the significance of an effect / difference.
•
• Formulas (with categorical variables) enable you to express rich links in your data.
• Visualizing your data and fitting simple models give insight into the data.

15.5. Testing for interactions 532 15.6. Full code for the figures 533
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

import pandas

import matplotlib.pyplot as plt

data = pandas.read_csv("brain_size.csv", sep=";", na_values=".")

# Box plot of FSIQ and PIQ (different measures od IQ)

plt.figure(figsize=(4, 3))
data.boxplot(column=["FSIQ", "PIQ"])

# Boxplot of the difference

plt.figure(figsize=(4, 3))
plt.boxplot(data["FSIQ"] - data["PIQ"])
plt.xticks((1,), ("FSIQ - PIQ",))

plt.show()

Total running time of the script: (0 minutes 0.074 seconds)

15.6.2 Plotting simple quantities of a pandas dataframe

This example loads from a CSV file data with mixed numerical and categorical entries, and plots a few
quantities, separately for females and males, thanks to the pandas integrated plotting tool (that uses
matplotlib behind the scene).
•
See http://pandas.pydata.org/pandas-docs/stable/visualization.html

•
•
import pandas

(continues on next page)

15.6. Full code for the figures 534 15.6. Full code for the figures 535
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

data = pandas.read_csv("brain_size.csv", sep=";", na_values=".")

# Box plots of different columns for each gender

groupby_gender = data.groupby("Gender")
groupby_gender.boxplot(column=["FSIQ", "VIQ", "PIQ"])

from pandas import plotting

# Scatter matrices for different columns

plotting.scatter_matrix(data[["Weight", "Height", "MRI_Count"]])
plotting.scatter_matrix(data[["PIQ", "VIQ", "FSIQ"]])

import matplotlib.pyplot as plt

plt.show()

Total running time of the script: (0 minutes 0.560 seconds)

15.6.3 Analysis of Iris petal and sepal sizes

Illustrate an analysis on a real dataset:
• Visualizing the data to formulate intuitions
• Fitting of a linear model
• Hypothesis test of the effect of a categorical variable in the presence of a continuous confound

import matplotlib.pyplot as plt Text(0.5, 0.98, 'blue: setosa, green: versicolor, red: virginica')

import pandas Statistical analysis

from pandas import plotting
# Let us try to explain the sepal length as a function of the petal
from statsmodels.formula.api import ols # width and the category of iris

# Load the data model = ols("sepal_width ~ name + petal_length", data).fit()

data = pandas.read_csv("iris.csv") print(model.summary())

Plot a scatter matrix # Now formulate a "contrast", to test if the offset for versicolor and
# virginica are identical
# Express the names as categories
categories = pandas.Categorical(data["name"]) print("Testing the difference between effect of versicolor and virginica")
print(model.f_test([0, 1, -1, 0]))
# The parameter 'c' is passed to plt.scatter and will control the color plt.show()
plotting.scatter_matrix(data, c=categories.codes, marker="o")

fig = plt.gcf() OLS Regression Results

fig.suptitle("blue: setosa, green: versicolor, red: virginica", size=13) ==============================================================================
Dep. Variable: sepal_width R-squared: 0.478
Model: OLS Adj. R-squared: 0.468
Method: Least Squares F-statistic: 44.63
Date: Thu, 10 Apr 2025 Prob (F-statistic): 1.58e-20
Time: 05:55:22 Log-Likelihood: -38.185
No. Observations: 150 AIC: 84.37
Df Residuals: 146 BIC: 96.41
Df Model: 3
Covariance Type: nonrobust
(continues on next page)

15.6. Full code for the figures 536 15.6. Full code for the figures 537
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

======================================================================================
coef std err t P>|t| [0.025 0.975]
--------------------------------------------------------------------------------------
Intercept 2.9813 0.099 29.989 0.000 2.785 3.178
name[T.versicolor] -1.4821 0.181 -8.190 0.000 -1.840 -1.124
name[T.virginica] -1.6635 0.256 -6.502 0.000 -2.169 -1.158
petal_length 0.2983 0.061 4.920 0.000 0.178 0.418
==============================================================================
Omnibus: 2.868 Durbin-Watson: 1.753
Prob(Omnibus): 0.238 Jarque-Bera (JB): 2.885
Skew: -0.082 Prob(JB): 0.236
Kurtosis: 3.659 Cond. No. 54.0
==============================================================================

Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly␣
˓→specified.

Testing the difference between effect of versicolor and virginica

<F test: F=3.245335346574177, p=0.0736905878170114, df_denom=146, df_num=1>

Total running time of the script: (0 minutes 0.440 seconds)

15.6.4 Simple Regression [<matplotlib.lines.Line2D object at 0x7f24628eaba0>]

Fit a simple linear regression using ‘statsmodels’, compute corresponding p-values.
Multilinear regression model, calculating fit, P-values, confidence intervals etc.
# Original author: Thomas Haslwanter
# Convert the data into a Pandas DataFrame to use the formulas framework
# in statsmodels
import numpy as np
data = pandas.DataFrame({"x": x, "y": y})
import matplotlib.pyplot as plt
import pandas
# Fit the model
model = ols("y ~ x", data).fit()
# For statistics. Requires statsmodels 5.0 or more
from statsmodels.formula.api import ols
# Print the summary
print(model.summary())
# Analysis of Variance (ANOVA) on linear models
from statsmodels.stats.anova import anova_lm
# Perform analysis of variance on fitted linear model
anova_results = anova_lm(model)
Generate and show the data

x = np.linspace(-5, 5, 20) print("\nANOVA results")

print(anova_results)
# To get reproducible values, provide a seed value
rng = np.random.default_rng(27446968) OLS Regression Results
==============================================================================
y = -5 + 3 * x + 4 * np.random.normal(size=x.shape) Dep. Variable: y R-squared: 0.845
Model: OLS Adj. R-squared: 0.836
# Plot the data Method: Least Squares F-statistic: 97.76
plt.figure(figsize=(5, 4)) Date: Thu, 10 Apr 2025 Prob (F-statistic): 1.06e-08
plt.plot(x, y, "o") Time: 05:55:22 Log-Likelihood: -53.560
No. Observations: 20 AIC: 111.1
Df Residuals: 18 BIC: 113.1
Df Model: 1
Covariance Type: nonrobust
==============================================================================
(continues on next page)

15.6. Full code for the figures 538 15.6. Full code for the figures 539
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
Intercept -4.1877 0.830 -5.044 0.000 -5.932 -2.444
x 2.7046 0.274 9.887 0.000 2.130 3.279
==============================================================================
Omnibus: 1.871 Durbin-Watson: 1.930
Prob(Omnibus): 0.392 Jarque-Bera (JB): 0.597
Skew: 0.337 Prob(JB): 0.742
Kurtosis: 3.512 Cond. No. 3.03
==============================================================================

Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly␣
˓→specified.

ANOVA results
df sum_sq mean_sq F PR(>F)
x 1.0 1347.476043 1347.476043 97.760281 1.062847e-08
Residual 18.0 248.102486 13.783471 NaN NaN

Plot the fitted model

# Retrieve the parameter estimates

offset, coef = model._results.params
plt.plot(x, x * coef + offset)
plt.xlabel("x")
plt.ylabel("y")

plt.show() Total running time of the script: (0 minutes 0.107 seconds)

15.6.5 Multiple Regression

Calculate using ‘statsmodels’ just the best fit, or all the corresponding statistical parameters.
Also shows how to make 3d plots.

# Original author: Thomas Haslwanter

import numpy as np
import matplotlib.pyplot as plt
import pandas

# For 3d plots. This import is necessary to have 3D plotting below

from mpl_toolkits.mplot3d import Axes3D

# For statistics. Requires statsmodels 5.0 or more

from statsmodels.formula.api import ols

# Analysis of Variance (ANOVA) on linear models

from statsmodels.stats.anova import anova_lm

Generate and show the data

x = np.linspace(-5, 5, 21)
# We generate a 2D grid
X, Y = np.meshgrid(x, x)
(continues on next page)

15.6. Full code for the figures 540 15.6. Full code for the figures 541
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

data = pandas.DataFrame({"x": X, "y": Y, "z": Z})
# To get reproducible values, provide a seed value
rng = np.random.default_rng(27446968) # Fit the model
model = ols("z ~ x + y", data).fit()
# Z is the elevation of this 2D grid
Z = -5 + 3 * X - 0.5 * Y + 8 * np.random.normal(size=X.shape) # Print the summary
print(model.summary())
# Plot the data
ax: Axes3D = plt.figure().add_subplot(projection="3d") print("\nRetrieving manually the parameter estimates:")
surf = ax.plot_surface(X, Y, Z, cmap="coolwarm", rstride=1, cstride=1) print(model._results.params)
ax.view_init(20, -120) # should be array([-4.99754526, 3.00250049, -0.50514907])
ax.set_xlabel("X")
ax.set_ylabel("Y") # Perform analysis of variance on fitted linear model
ax.set_zlabel("Z") anova_results = anova_lm(model)

print("\nANOVA results")
print(anova_results)

plt.show()

OLS Regression Results

==============================================================================
Dep. Variable: z R-squared: 0.579
Model: OLS Adj. R-squared: 0.577
Method: Least Squares F-statistic: 300.7
Date: Thu, 10 Apr 2025 Prob (F-statistic): 6.43e-83
Time: 05:55:23 Log-Likelihood: -1552.0
No. Observations: 441 AIC: 3110.
Df Residuals: 438 BIC: 3122.
Df Model: 2
Covariance Type: nonrobust
==============================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
Intercept -4.4332 0.390 -11.358 0.000 -5.200 -3.666
x 3.0861 0.129 23.940 0.000 2.833 3.340
y -0.6856 0.129 -5.318 0.000 -0.939 -0.432
==============================================================================
Omnibus: 0.560 Durbin-Watson: 1.967
Prob(Omnibus): 0.756 Jarque-Bera (JB): 0.651
Skew: -0.077 Prob(JB): 0.722
Kurtosis: 2.893 Cond. No. 3.03
==============================================================================

Text(-0.10764513121260137, 0.009865032686848034, 'Z') Notes:

[1] Standard Errors assume that the covariance matrix of the errors is correctly␣
Multilinear regression model, calculating fit, P-values, confidence intervals etc. ˓→specified.

# Convert the data into a Pandas DataFrame to use the formulas framework Retrieving manually the parameter estimates:
# in statsmodels [-4.43322435 3.08614608 -0.68556194]

# First we need to flatten the data: it's 2D layout is not relevant. ANOVA results
X = X.flatten() df sum_sq mean_sq F PR(>F)
Y = Y.flatten() x 1.0 38501.973182 38501.973182 573.111646 1.365553e-81
Z = Z.flatten() y 1.0 1899.955512 1899.955512 28.281320 1.676135e-07
Residual 438.0 29425.094352 67.180581 NaN NaN
(continues on next page)

15.6. Full code for the figures 542 15.6. Full code for the figures 543
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Total running time of the script: (0 minutes 0.114 seconds) (continued from previous page)
# Plot 2 linear fits for male and female.
seaborn.lmplot(y="wage", x="education", hue="gender", data=data)
15.6.6 Test for an education/gender interaction in wages
Wages depend mostly on education. Here we investigate how this dependence is related to gender: not
only does gender create an offset in wages, it also seems that wages increase more with education for
males than females.
Does our data support this last hypothesis? We will test this using statsmodels’ formulas (http://
statsmodels.sourceforge.net/stable/example_formulas.html).
Load and massage the data

import pandas

import urllib.request
import os

if not os.path.exists("wages.txt"):
# Download the file if it is not present
url = "http://lib.stat.cmu.edu/datasets/CPS_85_Wages"
with urllib.request.urlopen(url) as r, open("wages.txt", "wb") as f:
f.write(r.read())

# EDUCATION: Number of years of education

# SEX: 1=Female, 0=Male
# WAGE: Wage (dollars per hour)
data = pandas.read_csv(
"wages.txt",
skiprows=27,
skipfooter=6,
sep=None,
header=None,
names=["education", "gender", "wage"],
usecols=[0, 2, 5],
) <seaborn.axisgrid.FacetGrid object at 0x7f2462dadbe0>

# Convert genders to strings (this is particularly useful so that the statistical analysis
# statsmodels formulas detects that gender is a categorical variable)
import numpy as np import statsmodels.formula.api as sm

data["gender"] = np.choose(data.gender, ["male", "female"]) # Note that this model is not the plot displayed above: it is one
# joined model for male and female, not separate models for male and
# Log-transform the wages, because they typically are increased with # female. The reason is that a single model enables statistical testing
# multiplicative factors result = sm.ols(formula="wage ~ education + gender", data=data).fit()
data["wage"] = np.log10(data["wage"]) print(result.summary())

/home/runner/work/scientific-python-lectures/scientific-python-lectures/packages/ OLS Regression Results

˓→statistics/examples/plot_wage_education_gender.py:32: ParserWarning: Falling back␣ ==============================================================================
˓→to the 'python' engine because the 'c' engine does not support skipfooter; you can␣ Dep. Variable: wage R-squared: 0.193
˓→avoid this warning by specifying engine='python'. Model: OLS Adj. R-squared: 0.190
data = pandas.read_csv( Method: Least Squares F-statistic: 63.42
Date: Thu, 10 Apr 2025 Prob (F-statistic): 2.01e-25
simple plotting Time: 05:55:23 Log-Likelihood: 86.654
No. Observations: 534 AIC: -167.3
import seaborn Df Residuals: 531 BIC: -154.5
Df Model: 2
(continues on next page) (continues on next page)

15.6. Full code for the figures 544 15.6. Full code for the figures 545
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

Covariance Type: nonrobust ==============================================================================
==================================================================================
coef std err t P>|t| [0.025 0.975] Notes:
---------------------------------------------------------------------------------- [1] Standard Errors assume that the covariance matrix of the errors is correctly␣
Intercept 0.4053 0.046 8.732 0.000 0.314 0.496 ˓→specified.

gender[T.male] 0.1008 0.018 5.625 0.000 0.066 0.136

education 0.0334 0.003 9.768 0.000 0.027 0.040 Looking at the p-value of the interaction of gender and education, the data does not support the hy-
============================================================================== pothesis that education benefits males more than female (p-value > 0.05).
Omnibus: 4.675 Durbin-Watson: 1.792
Prob(Omnibus): 0.097 Jarque-Bera (JB): 4.876 import matplotlib.pyplot as plt
Skew: -0.147 Prob(JB): 0.0873
Kurtosis: 3.365 Cond. No. 69.7 plt.show()
==============================================================================
Total running time of the script: (0 minutes 0.520 seconds)
Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly␣
˓→specified.
15.6.7 Visualizing factors influencing wages
The plots above highlight that there is not only a different offset in wage but also a different slope This example uses seaborn to quickly plot various factors relating wages, experience, and education.
We need to model this using an interaction Seaborn (https://seaborn.pydata.org) is a library that combines visualization and statistical fits to show
trends in data.
result = sm.ols(
formula="wage ~ education + gender + education * gender", data=data Note that importing seaborn changes the matplotlib style to have an “excel-like” feeling. This changes
).fit() affect other matplotlib figures. To restore defaults once this example is run, we would need to call
print(result.summary()) plt.rcdefaults().

# Standard library imports

OLS Regression Results import os
==============================================================================
Dep. Variable: wage R-squared: 0.198 import matplotlib.pyplot as plt
Model: OLS Adj. R-squared: 0.194
Method: Least Squares F-statistic: 43.72
Load the data
Date: Thu, 10 Apr 2025 Prob (F-statistic): 2.94e-25
Time: 05:55:23 Log-Likelihood: 88.503 import pandas
No. Observations: 534 AIC: -169.0 import requests
Df Residuals: 530 BIC: -151.9
Df Model: 3 if not os.path.exists("wages.txt"):
Covariance Type: nonrobust # Download the file if it is not present
============================================================================================ r = requests.get("http://lib.stat.cmu.edu/datasets/CPS_85_Wages")
coef std err t P>|t| [0.025 ␣ with open("wages.txt", "wb") as f:
˓→0.975] f.write(r.content)
--------------------------------------------------------------------------------------
˓→------ # Give names to the columns
Intercept 0.2998 0.072 4.173 0.000 0.159 ␣ names = [
˓→ 0.441 "EDUCATION: Number of years of education",
gender[T.male] 0.2750 0.093 2.972 0.003 0.093 ␣ "SOUTH: 1=Person lives in South, 0=Person lives elsewhere",
˓→ 0.457 "SEX: 1=Female, 0=Male",
education 0.0415 0.005 7.647 0.000 0.031 ␣ "EXPERIENCE: Number of years of work experience",
˓→ 0.052 "UNION: 1=Union member, 0=Not union member",
education:gender[T.male] -0.0134 0.007 -1.919 0.056 -0.027 ␣ "WAGE: Wage (dollars per hour)",
˓→ 0.000 "AGE: years",
============================================================================== "RACE: 1=Other, 2=Hispanic, 3=White",
Omnibus: 4.838 Durbin-Watson: 1.825 "OCCUPATION: 1=Management, 2=Sales, 3=Clerical, 4=Service, 5=Professional, 6=Other
Prob(Omnibus): 0.089 Jarque-Bera (JB): 5.000 ˓→",
Skew: -0.156 Prob(JB): 0.0821 "SECTOR: 0=Other, 1=Manufacturing, 2=Construction",
Kurtosis: 3.356 Cond. No. 194. "MARR: 0=Unmarried, 1=Married",
(continues on next page) (continues on next page)

15.6. Full code for the figures 546 15.6. Full code for the figures 547
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

]

short_names = [n.split(":")[0] for n in names]

data = pandas.read_csv(
"wages.txt", skiprows=27, skipfooter=6, sep=None, header=None, engine="python"
)
data.columns = pandas.Index(short_names)

# Log-transform the wages, because they typically are increased with

# multiplicative factors
import numpy as np

data["WAGE"] = np.log10(data["WAGE"])

Plot scatter matrices highlighting different aspects

import seaborn

seaborn.pairplot(data, vars=["WAGE", "AGE", "EDUCATION"], kind="reg")

seaborn.pairplot(data, vars=["WAGE", "AGE", "EDUCATION"], kind="reg", hue="SEX")

plt.suptitle("Effect of gender: 1=Female, 0=Male")

seaborn.pairplot(data, vars=["WAGE", "AGE", "EDUCATION"], kind="reg", hue="RACE")

plt.suptitle("Effect of race: 1=Other, 2=Hispanic, 3=White")

seaborn.pairplot(data, vars=["WAGE", "AGE", "EDUCATION"], kind="reg", hue="UNION")

plt.suptitle("Effect of union: 1=Union member, 0=Not union member")

15.6. Full code for the figures 548 15.6. Full code for the figures 549
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• •

15.6. Full code for the figures 550 15.6. Full code for the figures 551
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Total running time of the script: (0 minutes 8.422 seconds)

•

Text(0.5, 0.98, 'Effect of union: 1=Union member, 0=Not union member')

15.6.8 Air fares before and after 9/11
Plot a simple regression This is a business-intelligence (BI) like application.

seaborn.lmplot(y="WAGE", x="EDUCATION", data=data) What is interesting here is that we may want to study fares as a function of the year, paired accordingly
to the trips, or forgetting the year, only as a function of the trip endpoints.
plt.show() Using statsmodels’ linear models, we find that both with an OLS (ordinary least square) and a robust
fit, the intercept and the slope are significantly non-zero: the air fares have decreased between 2000 and
2001, and their dependence on distance travelled has also decreased
# Standard library imports
import os

Load the data

import pandas
import requests

if not os.path.exists("airfares.txt"):
# Download the file if it is not present
r = requests.get(
"https://users.stat.ufl.edu/~winner/data/airq4.dat",
verify=False, # Wouldn't normally do this, but this site's certificate
# is not yet distributed
(continues on next page)

15.6. Full code for the figures 552 15.6. Full code for the figures 553
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

) # Rename the columns
with open("airfares.txt", "wb") as f: data_2001.columns = pandas.Index(
f.write(r.content) ["city1", "city2", "pop1", "pop2", "dist", "fare", "nb_passengers"]
)
# As a separator, ' +' is a regular expression that means 'one of more # Add a column with the year
# space' data_2001.insert(0, "year", 2001)
data = pandas.read_csv(
"airfares.txt", data_flat = pandas.concat([data_2000, data_2001])
delim_whitespace=True,
header=0, Plot scatter matrices highlighting different aspects
names=[
"city1", import seaborn
"city2",
"pop1", seaborn.pairplot(
"pop2", data_flat, vars=["fare", "dist", "nb_passengers"], kind="reg", markers="."
"dist", )
"fare_2000",
"nb_passengers_2000", # A second plot, to show the effect of the year (ie the 9/11 effect)
"fare_2001", seaborn.pairplot(
"nb_passengers_2001", data_flat,
], vars=["fare", "dist", "nb_passengers"],
) kind="reg",
hue="year",
# we log-transform the number of passengers markers=".",
import numpy as np )

data["nb_passengers_2000"] = np.log10(data["nb_passengers_2000"])
data["nb_passengers_2001"] = np.log10(data["nb_passengers_2001"])

/home/runner/work/scientific-python-lectures/scientific-python-lectures/packages/
˓→statistics/examples/plot_airfare.py:38: FutureWarning: The 'delim_whitespace'␣

˓→keyword in pd.read_csv is deprecated and will be removed in a future version. Use␣

˓→``sep='\s+'`` instead

data = pandas.read_csv(

Make a dataframe with the year as an attribute, instead of separate columns

# This involves a small danse in which we separate the dataframes in 2,

# one for year 2000, and one for 2001, before concatenating again.

# Make an index of each flight

data_flat = data.reset_index()

data_2000 = data_flat[
["city1", "city2", "pop1", "pop2", "dist", "fare_2000", "nb_passengers_2000"]
]
# Rename the columns
data_2000.columns = pandas.Index(
["city1", "city2", "pop1", "pop2", "dist", "fare", "nb_passengers"]
)
# Add a column with the year
data_2000.insert(0, "year", 2000)

data_2001 = data_flat[
["city1", "city2", "pop1", "pop2", "dist", "fare_2001", "nb_passengers_2001"]
]
(continues on next page)

15.6. Full code for the figures 554 15.6. Full code for the figures 555
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

<seaborn.axisgrid.PairGrid object at 0x7f24639c16a0>

• Plot the difference in fare

import matplotlib.pyplot as plt

plt.figure(figsize=(5, 2))
seaborn.boxplot(data.fare_2001 - data.fare_2000)
plt.title("Fare: 2001 - 2000")
plt.subplots_adjust()

plt.figure(figsize=(5, 2))
seaborn.boxplot(data.nb_passengers_2001 - data.nb_passengers_2000)
plt.title("NB passengers: 2001 - 2000")
plt.subplots_adjust()

15.6. Full code for the figures 556 15.6. Full code for the figures 557
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

Kurtosis: 3.338 Cond. No. 5.23e+03
==============================================================================

Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly␣
˓→specified.

[2] The condition number is large, 5.23e+03. This might indicate that there are
strong multicollinearity or other numerical problems.
Robust linear Model Regression Results
==============================================================================
• Dep. Variable: fare No. Observations: 8352
Model: RLM Df Residuals: 8349
Method: IRLS Df Model: 2
Norm: HuberT
Scale Est.: mad
Cov Type: H1
Date: Thu, 10 Apr 2025
Time: 05:55:39
No. Iterations: 12
=================================================================================
coef std err z P>|z| [0.025 0.975]
---------------------------------------------------------------------------------
Intercept 215.0848 2.448 87.856 0.000 210.287 219.883
• dist 0.0460 0.001 46.166 0.000 0.044 0.048
Statistical testing: dependence of fare on distance and number of passengers nb_passengers -35.2686 1.119 -31.526 0.000 -37.461 -33.076
=================================================================================
import statsmodels.formula.api as sm
If the model instance has been used for another fit with different fit parameters,␣
result = sm.ols(formula="fare ~ 1 + dist + nb_passengers", data=data_flat).fit() ˓→then the fit options might not be the correct ones anymore .

print(result.summary())
Statistical testing: regression of fare on distance: 2001/2000 difference
# Using a robust fit
result = sm.rlm(formula="fare ~ 1 + dist + nb_passengers", data=data_flat).fit() result = sm.ols(formula="fare_2001 - fare_2000 ~ 1 + dist", data=data).fit()
print(result.summary()) print(result.summary())

OLS Regression Results # Plot the corresponding regression

============================================================================== data["fare_difference"] = data["fare_2001"] - data["fare_2000"]
Dep. Variable: fare R-squared: 0.275 seaborn.lmplot(x="dist", y="fare_difference", data=data)
Model: OLS Adj. R-squared: 0.275
Method: Least Squares F-statistic: 1585. plt.show()
Date: Thu, 10 Apr 2025 Prob (F-statistic): 0.00
Time: 05:55:39 Log-Likelihood: -45532.
No. Observations: 8352 AIC: 9.107e+04
Df Residuals: 8349 BIC: 9.109e+04
Df Model: 2
Covariance Type: nonrobust
=================================================================================
coef std err t P>|t| [0.025 0.975]
---------------------------------------------------------------------------------
Intercept 211.2428 2.466 85.669 0.000 206.409 216.076
dist 0.0484 0.001 48.149 0.000 0.046 0.050
nb_passengers -32.8925 1.127 -29.191 0.000 -35.101 -30.684
==============================================================================
Omnibus: 604.051 Durbin-Watson: 1.446
Prob(Omnibus): 0.000 Jarque-Bera (JB): 740.733
Skew: 0.710 Prob(JB): 1.42e-161
(continues on next page)

15.6. Full code for the figures 558 15.6. Full code for the figures 559
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

[2] The condition number is large, 2.4e+03. This might indicate that there are
strong multicollinearity or other numerical problems.

Total running time of the script: (0 minutes 7.097 seconds)

15.7 Solutions to this chapter’s exercises

15.7.1 Solutions to this chapter’s exercises
Relating Gender and IQ
Going back to the brain size + IQ data, test if the VIQ of male and female are different after removing
the effect of brain size, height and weight.
Notice that here ‘Gender’ is a categorical value. As it is a non-float data type, statsmodels is able to
automatically infer this.

import pandas
from statsmodels.formula.api import ols

data = pandas.read_csv("../brain_size.csv", sep=";", na_values=".")

model = ols("VIQ ~ Gender + MRI_Count + Height", data).fit()

print(model.summary())

# Here, we don't need to define a contrast, as we are testing a single

# coefficient of our model, and not a combination of coefficients.
# However, defining a contrast, which would then be a 'unit contrast',
# will give us the same results
OLS Regression Results print(model.f_test([0, 1, 0, 0]))
==============================================================================
Dep. Variable: fare_2001 R-squared: 0.159 OLS Regression Results
Model: OLS Adj. R-squared: 0.159 ==============================================================================
Method: Least Squares F-statistic: 791.7 Dep. Variable: VIQ R-squared: 0.246
Date: Thu, 10 Apr 2025 Prob (F-statistic): 1.20e-159 Model: OLS Adj. R-squared: 0.181
Time: 05:55:39 Log-Likelihood: -22640. Method: Least Squares F-statistic: 3.809
No. Observations: 4176 AIC: 4.528e+04 Date: Thu, 10 Apr 2025 Prob (F-statistic): 0.0184
Df Residuals: 4174 BIC: 4.530e+04 Time: 05:55:39 Log-Likelihood: -172.34
Df Model: 1 No. Observations: 39 AIC: 352.7
Covariance Type: nonrobust Df Residuals: 35 BIC: 359.3
============================================================================== Df Model: 3
coef std err t P>|t| [0.025 0.975] Covariance Type: nonrobust
------------------------------------------------------------------------------ ==================================================================================
Intercept 148.0279 1.673 88.480 0.000 144.748 151.308 coef std err t P>|t| [0.025 0.975]
dist 0.0388 0.001 28.136 0.000 0.036 0.041 ----------------------------------------------------------------------------------
============================================================================== Intercept 166.6258 88.824 1.876 0.069 -13.696 346.948
Omnibus: 136.558 Durbin-Watson: 1.544 Gender[T.Male] 8.8524 10.710 0.827 0.414 -12.890 30.595
Prob(Omnibus): 0.000 Jarque-Bera (JB): 149.624 MRI_Count 0.0002 6.46e-05 2.615 0.013 3.78e-05 0.000
Skew: 0.462 Prob(JB): 3.23e-33 Height -3.0837 1.276 -2.417 0.021 -5.674 -0.494
Kurtosis: 2.920 Cond. No. 2.40e+03 ==============================================================================
============================================================================== Omnibus: 7.373 Durbin-Watson: 2.109
Prob(Omnibus): 0.025 Jarque-Bera (JB): 2.252
Notes: Skew: 0.005 Prob(JB): 0.324
[1] Standard Errors assume that the covariance matrix of the errors is correctly␣ Kurtosis: 1.823 Cond. No. 2.40e+07
˓→specified.
==============================================================================
(continues on next page) (continues on next page)

15.6. Full code for the figures 560 15.7. Solutions to this chapter’s exercises 561
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly␣
˓→specified.

[2] The condition number is large, 2.4e+07. This might indicate that there are
strong multicollinearity or other numerical problems.
<F test: F=0.683196084584229, p=0.4140878441244694, df_denom=35, df_num=1>

Here we plot a scatter matrix to get intuitions on our results. This goes beyond what was asked in the
exercise

# This plotting is useful to get an intuitions on the relationships between

# our different variables

from pandas import plotting

import matplotlib.pyplot as plt

# Fill in the missing values for Height for plotting

data["Height"] = data["Height"].ffill()

# The parameter 'c' is passed to plt.scatter and will control the color
# The same holds for parameters 'marker', 'alpha' and 'cmap', that
# control respectively the type of marker used, their transparency and
# the colormap
plotting.scatter_matrix(
data[["VIQ", "MRI_Count", "Height"]],
c=(data["Gender"] == "Female"),
marker="o",
alpha=1, Total running time of the script: (0 minutes 0.232 seconds)
cmap="winter",
)

fig = plt.gcf()
fig.suptitle("blue: male, green: female", size=13)

plt.show()

15.7. Solutions to this chapter’s exercises 562 15.7. Solutions to this chapter’s exercises 563
 Scientific Python Lectures, Edition 2025.1rc0.dev0

• Algebraic manipulations
– Expand
– Simplify

16
• Calculus
– Limits
– Differentiation
– Series expansion

CHAPTER – Integration
• Equation solving
• Linear Algebra
– Matrices
– Differential Equations

Sympy : Symbolic Mathematics in 16.1 First Steps with SymPy

Python 16.1.1 Using SymPy as a calculator
SymPy defines three numerical types: Real, Rational and Integer.
The Rational class represents a rational number as a pair of two Integers: the numerator and the
denominator, so Rational(1, 2) represents 1/2, Rational(5, 2) 5/2 and so on:

>>> import sympy as sym

>>> a = sym.Rational(1, 2)
Author: Fabian Pedregosa
>>> a
1/2
Objectives
>>> a*2
1. Evaluate expressions with arbitrary precision.
1
2. Perform algebraic manipulations on symbolic expressions.
3. Perform basic calculus tasks (limits, differentiation and SymPy uses mpmath in the background, which makes it possible to perform computations using
integration) with symbolic expressions. arbitrary-precision arithmetic. That way, some special constants, like 𝑒, 𝑝𝑖, 𝑜𝑜 (Infinity), are treated
as symbols and can be evaluated with arbitrary precision:
4. Solve polynomial and transcendental equations.
>>> sym.pi**2
5. Solve some differential equations. pi**2

What is SymPy? SymPy is a Python library for symbolic mathematics. It aims to be an alternative to >>> sym.pi.evalf()
systems such as Mathematica or Maple while keeping the code as simple as possible and easily extensible. 3.14159265358979
SymPy is written entirely in Python and does not require any external libraries.
Sympy documentation and packages for installation can be found on https://www.sympy.org/ >>> (sym.pi + sym.exp(1)).evalf()
5.85987448204884

Chapters contents as you see, evalf evaluates the expression to a floating-point number.
There is also a class representing mathematical infinity, called oo:
• First Steps with SymPy
– Using SymPy as a calculator >>> sym.oo > 99999
True
– Symbols (continues on next page)

564 16.1. First Steps with SymPy 565

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> sym.expand(x + y, complex=True)
>>> sym.oo + 1 re(x) + re(y) + I*im(x) + I*im(y)
oo >>> sym.I * sym.im(x) + sym.I * sym.im(y) + sym.re(x) + sym.re(y)
re(x) + re(y) + I*im(x) + I*im(y)

Exercises >>> sym.expand(sym.cos(x + y), trig=True)

√ -sin(x)*sin(y) + cos(x)*cos(y)
1. Calculate 2 with 100 decimals. >>> sym.cos(x) * sym.cos(y) - sym.sin(x) * sym.sin(y)
2. Calculate 1/2 + 1/3 in rational arithmetic. -sin(x)*sin(y) + cos(x)*cos(y)

16.2.2 Simplify
16.1.2 Symbols
Use simplify if you would like to transform an expression into a simpler form:
In contrast to other Computer Algebra Systems, in SymPy you have to declare symbolic variables
explicitly: >>> sym.simplify((x + x * y) / x)
y + 1
>>> x = sym.Symbol('x')
>>> y = sym.Symbol('y')
Simplification is a somewhat vague term, and more precises alternatives to simplify exists: powsimp
(simplification of exponents), trigsimp (for trigonometric expressions) , logcombine, radsimp, together.
Then you can manipulate them:

>>> x + y + x - y Exercises
2*x
1. Calculate the expanded form of (𝑥 + 𝑦)6 .
>>> (x + y) ** 2
(x + y)**2 2. Simplify the trigonometric expression sin(𝑥)/ cos(𝑥)

Symbols can now be manipulated using some of python operators: +, -, *, ** (arithmetic), &, |, ~, >>,
<< (boolean). 16.3 Calculus

Printing
16.3.1 Limits
Limits are easy to use in SymPy, they follow the syntax limit(function, variable, point), so to
Sympy allows for control of the display of the output. From here we use the following setting for compute the limit of 𝑓 (𝑥) as 𝑥 → 0, you would issue limit(f, x, 0):
printing:
>>> sym.limit(sym.sin(x) / x, x, 0)
>>> sym.init_printing(use_unicode=False, wrap_line=True)
1

you can also calculate the limit at infinity:

16.2 Algebraic manipulations >>> sym.limit(x, x, sym.oo)
oo
SymPy is capable of performing powerful algebraic manipulations. We’ll take a look into some of the
most frequently used: expand and simplify.
>>> sym.limit(1 / x, x, sym.oo)
0
16.2.1 Expand
Use this to expand an algebraic expression. It will try to denest powers and multiplications: >>> sym.limit(x ** x, x, 0)
1
>>> sym.expand((x + y) ** 3)
3 2 2 3
x + 3*x *y + 3*x*y + y 16.3.2 Differentiation
>>> 3 * x * y ** 2 + 3 * y * x ** 2 + x ** 3 + y ** 3 You can differentiate any SymPy expression using diff(func, var). Examples:
3 2 2 3
x + 3*x *y + 3*x*y + y >>> sym.diff(sym.sin(x), x)
cos(x)
Further options can be given in form on keywords: >>> sym.diff(sym.sin(2 * x), x)
2*cos(2*x)

(continues on next page)

16.2. Algebraic manipulations 566 16.3. Calculus 567

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) 16.3.4 Integration

>>> sym.diff(sym.tan(x), x) SymPy has support for indefinite and definite integration of transcendental elementary and special func-
2 tions via integrate() facility, which uses the powerful extended Risch-Norman algorithm and some
tan (x) + 1 heuristics and pattern matching. You can integrate elementary functions:

You can check that it is correct by: >>> sym.integrate(6 * x ** 5, x)

6
>>> sym.limit((sym.tan(x + y) - sym.tan(x)) / y, y, 0) x
1 >>> sym.integrate(sym.sin(x), x)
------- -cos(x)
2 >>> sym.integrate(sym.log(x), x)
cos (x) x*log(x) - x
>>> sym.integrate(2 * x + sym.sinh(x), x)
Which is equivalent since 2
x + cosh(x)
1
sec(𝑥) = 𝑎𝑛𝑑 sec2 (𝑥) = tan2 (𝑥) + 1.
cos(𝑥)
Also special functions are handled easily:
You can check this as well:
>>> sym.integrate(sym.exp(-x ** 2) * sym.erf(x), x)
>>> sym.trigsimp(sym.diff(sym.tan(x), x)) ____ 2
1 \/ pi *erf (x)
------- --------------
2 4
cos (x)
It is possible to compute definite integral:
Higher derivatives can be calculated using the diff(func, var, n) method:
>>> sym.integrate(x**3, (x, -1, 1))
>>> sym.diff(sym.sin(2 * x), x, 1) 0
2*cos(2*x) >>> sym.integrate(sym.sin(x), (x, 0, sym.pi / 2))
1
>>> sym.diff(sym.sin(2 * x), x, 2) >>> sym.integrate(sym.cos(x), (x, -sym.pi / 2, sym.pi / 2))
-4*sin(2*x) 2

>>> sym.diff(sym.sin(2 * x), x, 3) Also improper integrals are supported as well:

-8*cos(2*x)
>>> sym.integrate(sym.exp(-x), (x, 0, sym.oo))
1
16.3.3 Series expansion >>> sym.integrate(sym.exp(-x ** 2), (x, -sym.oo, sym.oo))
____
SymPy also knows how to compute the Taylor series of an expression at a point. Use series(expr, \/ pi
var):

>>> sym.series(sym.cos(x), x)
2 4 16.4 Equation solving
x x / 6\
1 - -- + -- + O\x / SymPy is able to solve algebraic equations, in one and several variables using solveset():
2 24
>>> sym.solveset(x ** 4 - 1, x)
>>> sym.series(1/sym.cos(x), x)
{-1, 1, -I, I}
2 4
x 5*x / 6\
As you can see it takes as first argument an expression that is supposed to be equaled to 0. It also has
1 + -- + ---- + O\x /
(limited) support for transcendental equations:
2 24
>>> sym.solveset(sym.exp(x) + 1, x)
{I*(2*n*pi + pi) | n in Integers}
Exercises

1. Calculate lim𝑥→0 sin(𝑥)/𝑥

Systems of linear equations
2. Calculate the derivative of 𝑙𝑜𝑔(𝑥) for 𝑥.

16.3. Calculus 568 16.4. Equation solving 569

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Sympy is able to solve a large part of polynomial equations, and is also capable of solving multiple (continued from previous page)
equations with respect to multiple variables giving a tuple as second argument. To do this you use
the solve() command: >>> A**2
>>> solution = sym.solve((x + 5 * y - 2, -3 * x + 6 * y - 15), (x, y)) [x*y + 1 2*x ]
>>> solution[x], solution[y] [ ]
(-3, 1) [ 2*y x*y + 1]

Another alternative in the case of polynomial equations is factor. factor returns the polynomial factorized
into irreducible terms, and is capable of computing the factorization over various domains:
16.5.2 Differential Equations
SymPy is capable of solving (some) Ordinary Differential. To solve differential equations, use dsolve.
>>> f = x ** 4 - 3 * x ** 2 + 1 First, create an undefined function by passing cls=Function to the symbols function:
>>> sym.factor(f)
/ 2 \ / 2 \ >>> f, g = sym.symbols('f g', cls=sym.Function)
\x - x - 1/*\x + x - 1/
f and g are now undefined functions. We can call f(x), and it will represent an unknown function:
>>> sym.factor(f, modulus=5)
2 2 >>> f(x)
(x - 2) *(x + 2) f(x)

SymPy is also able to solve boolean equations, that is, to decide if a certain boolean expression is >>> f(x).diff(x, x) + f(x)
satisfiable or not. For this, we use the function satisfiable: 2
d
>>> sym.satisfiable(x & y) f(x) + ---(f(x))
{x: True, y: True} 2
dx
This tells us that (x & y) is True whenever x and y are both True. If an expression cannot be true, i.e.
no values of its arguments can make the expression True, it will return False: >>> sym.dsolve(f(x).diff(x, x) + f(x), f(x))
f(x) = C1*sin(x) + C2*cos(x)
>>> sym.satisfiable(x & ~x)
False Keyword arguments can be given to this function in order to help if find the best possible resolution sys-
tem. For example, if you know that it is a separable equations, you can use keyword hint='separable'
to force dsolve to resolve it as a separable equation:
Exercises
>>> sym.dsolve(sym.sin(x) * sym.cos(f(x)) + sym.cos(x) * sym.sin(f(x)) * f(x).diff(x),
1. Solve the system of equations 𝑥 + 𝑦 = 2, 2 · 𝑥 + 𝑦 = 0 ˓→ f(x), hint='separable')

2. Are there boolean values x, y that make (~x | y) & (~y | x) true? / C1 \ / C1 \
[f(x) = - acos|------| + 2*pi, f(x) = acos|------|]
\cos(x)/ \cos(x)/

16.5 Linear Algebra

Exercises
16.5.1 Matrices
1. Solve the Bernoulli differential equation
Matrices are created as instances from the Matrix class:

>>> sym.Matrix([[1, 0], [0, 1]])

[1 0] 𝑑𝑓 (𝑥)
𝑥 + 𝑓 (𝑥) − 𝑓 (𝑥)2 = 0
[ ] 𝑥
[0 1] 2. Solve the same equation using hint='Bernoulli'. What do you observe ?

unlike a NumPy array, you can also put Symbols in it:

>>> x, y = sym.symbols('x, y')

>>> A = sym.Matrix([[1, x], [y, 1]])
>>> A
[1 x]
[ ]
[y 1]
(continues on next page)

16.5. Linear Algebra 570 16.5. Linear Algebra 571

 Scientific Python Lectures, Edition 2025.1rc0.dev0

– Colorspaces
• Image preprocessing / enhancement
– Local filters

17
– Non-local filters
– Mathematical morphology
• Image segmentation
– Binary segmentation: foreground + background

CHAPTER – Marker based methods

• Measuring regions’ properties
• Data visualization and interaction
• Feature extraction for computer vision
• Full code examples
• Examples for the scikit-image chapter

scikit-image: image processing

17.1 Introduction and concepts
Images are NumPy’s arrays np.ndarray
image
np.ndarray
Author: Emmanuelle Gouillart pixels
array values: a[2, 3]
scikit-image is a Python package dedicated to image processing, using NumPy arrays as image objects.
This chapter describes how to use scikit-image for various image processing tasks, and how it relates channels
to other scientific Python modules such as NumPy and SciPy. array dimensions
image encoding
ã See also dtype (np.uint8, np.uint16, np.float)
filters
For basic image manipulation, such as image cropping or simple filtering, a large number of simple
functions (numpy, skimage, scipy)
operations can be realized with NumPy and SciPy only. See Image manipulation and processing using
NumPy and SciPy. >>> import numpy as np
Note that you should be familiar with the content of the previous chapter before reading the current >>> check = np.zeros((8, 8))
one, as basic operations such as masking and labeling are a prerequisite. >>> check[::2, 1::2] = 1
>>> check[1::2, ::2] = 1
>>> import matplotlib.pyplot as plt
>>> plt.imshow(check, cmap='gray', interpolation='nearest')
Chapters contents <matplotlib.image.AxesImage object at ...>
• Introduction and concepts
– scikit-image and the scientific Python ecosystem
– What is included in scikit-image
• Importing
• Example data
• Input/output, data types and colorspaces
– Data types

572 17.1. Introduction and concepts 573

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

feature
Feature detection and extraction, e.g., texture analysis corners, etc.
filters
Sharpening, edge finding, rank filters, thresholding, etc.
graph
Graph-theoretic operations, e.g., shortest paths.
io
Reading, saving, and displaying images and video.
measure
Measurement of image properties, e.g., region properties and contours.
metrics
Metrics corresponding to images, e.g. distance metrics, similarity, etc.
morphology
Morphological operations, e.g., opening or skeletonization.
restoration
17.1.1 scikit-image and the scientific Python ecosystem Restoration algorithms, e.g., deconvolution algorithms, denoising, etc.

scikit-image is packaged in both pip and conda-based Python installations, as well as in most Linux segmentation
distributions. Other Python packages for image processing & visualization that operate on NumPy Partitioning an image into multiple regions.
arrays include: transform
scipy.ndimage Geometric and other transforms, e.g., rotation or the Radon transform.
For N-dimensional arrays. Basic filtering, mathematical morphology, regions properties util
Mahotas Generic utilities.
With a focus on high-speed implementations.
Napari 17.2 Importing
A fast, interactive, multi-dimensional image viewer built in Qt.
We import scikit-image using the convention:
Some powerful C++ image processing libraries also have Python bindings:
>>> import skimage as ski
OpenCV
A highly optimized computer vision library with a focus on real-time applications.
Most functionality lives in subpackages, e.g.:
ITK
The Insight ToolKit, especially useful for registration and working with 3D images. >>> image = ski.data.cat()
To varying degrees, these tend to be less Pythonic and NumPy-friendly.
You can list all submodules with:

17.1.2 What is included in scikit-image >>> for m in dir(ski): print(m)

__version__
• Website: https://scikit-image.org/ color
• Gallery of examples: https://scikit-image.org/docs/stable/auto_examples/ data
draw
The library contains predominantly image processing algorithms, but also utility functions to ease data exposure
handling and processing. It contains the following submodules: feature
color filters
Color space conversion. future
graph
data io
Test images and example data. measure
draw metrics
Drawing primitives (lines, text, etc.) that operate on NumPy arrays. morphology
registration
exposure restoration
Image intensity adjustment, e.g., histogram equalization, etc. segmentation
(continues on next page)

17.1. Introduction and concepts 574 17.2. Importing 575

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) Loading also works with URLs:

transform
>>> logo = ski.io.imread('https://scikit-image.org/_static/img/logo.png')
util

Most scikit-image functions take NumPy ndarrays as arguments 17.4.1 Data types
>>> camera = ski.data.camera()
>>> camera.dtype
dtype('uint8')
>>> camera.shape
(512, 512)
>>> filtered_camera = ski.filters.gaussian(camera, sigma=1)
>>> type(filtered_camera)
<class 'numpy.ndarray'>

17.3 Example data Image ndarrays can be represented either by integers (signed or unsigned) or floats.

To start off, we need example images to work with. The library ships with a few of these: Careful with overflows with integer data types

skimage.data >>> camera = ski.data.camera()

>>> camera.dtype
>>> image = ski.data.cat() dtype('uint8')
>>> image.shape >>> camera_multiply = 3 * camera
(300, 451, 3)
Different integer sizes are possible: 8-, 16- or 32-bytes, signed or unsigned.

17.4 Input/output, data types and colorspaces . Warning

I/O: skimage.io An important (if questionable) skimage convention: float images are supposed to lie in [-1, 1] (in
Save an image to disk: skimage.io.imsave() order to have comparable contrast for all float images)

>>> ski.io.imsave("cat.png", image) >>> camera_float = ski.util.img_as_float(camera)

>>> camera.max(), camera_float.max()
(np.uint8(255), np.float64(1.0))
Reading from files: skimage.io.imread()

>>> cat = ski.io.imread("cat.png")

Some image processing routines need to work with float arrays, and may hence output an array with a
different type and the data range from the input array
>>> camera_sobel = ski.filters.sobel(camera)
>>> camera_sobel.max()
np.float64(0.644...)

Utility functions are provided in skimage to convert both the dtype and the data range, following
skimage’s conventions: util.img_as_float, util.img_as_ubyte, etc.
See the user guide for more details.

17.4.2 Colorspaces
Color images are of shape (N, M, 3) or (N, M, 4) (when an alpha channel encodes transparency)
>>> face = sp.datasets.face()
>>> face.shape
(768, 1024, 3)

Routines converting between different colorspaces (RGB, HSV, LAB etc.) are available in skimage.
color : color.rgb2hsv, color.lab2rgb, etc. Check the docstring for the expected dtype (and data
This works with many data formats supported by the ImageIO library. range) of input images.

17.3. Example data 576 17.4. Input/output, data types and colorspaces 577
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

17.5.2 Non-local filters

3D images Non-local filters use a large region of the image (or all the image) to transform the value of one pixel:
Most functions of skimage can take 3D images as input arguments. Check the docstring to know if >>> camera = ski.data.camera()
a function can be used on 3D images (for example MRI or CT images). >>> camera_equalized = ski.exposure.equalize_hist(camera)

Enhances contrast in large almost uniform regions.

Exercise

Open a color image on your disk as a NumPy array.

Find a skimage function computing the histogram of an image and plot the histogram of
each color channel
Convert the image to grayscale and plot its histogram.

17.5 Image preprocessing / enhancement

Goals: denoising, feature (edges) extraction, . . .

17.5.1 Local filters 17.5.3 Mathematical morphology

Local filters replace the value of pixels by a function of the values of neighboring pixels. The function See wikipedia for an introduction on mathematical morphology.
can be linear or non-linear.
Probe an image with a simple shape (a structuring element), and modify this image according to how
Neighbourhood: square (choose size), disk, or more complicated structuring element. the shape locally fits or misses the image.
Default structuring element: 4-connectivity of a pixel

>>> # Import structuring elements to make them more easily accessible

>>> from skimage.morphology import disk, diamond

>>> diamond(1)
array([[0, 1, 0],
[1, 1, 1],
[0, 1, 0]], dtype=uint8)

Example : horizontal Sobel filter

>>> text = ski.data.text()

>>> hsobel_text = ski.filters.sobel_h(text)

Uses the following linear kernel for computing horizontal gradients:

1 2 1
0 0 0
-1 -2 -1

Erosion = minimum filter. Replace the value of a pixel by the minimal value covered by the structuring
element.:

>>> a = np.zeros((7,7), dtype=np.uint8)

>>> a[1:6, 2:5] = 1
>>> a
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
(continues on next page)

17.5. Image preprocessing / enhancement 578 17.5. Image preprocessing / enhancement 579
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

[0, 0, 1, 1, 1, 0, 0], Grayscale mathematical morphology
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0], Mathematical morphology operations are also available for (non-binary) grayscale images (int or float
[0, 0, 0, 0, 0, 0, 0]], dtype=uint8) type). Erosion and dilation correspond to minimum (resp. maximum) filters.
>>> ski.morphology.binary_erosion(a, diamond(1)).astype(np.uint8)
array([[0, 0, 0, 0, 0, 0, 0], Higher-level mathematical morphology are available: tophat, skeletonization, etc.
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0], ã See also
[0, 0, 0, 1, 0, 0, 0], Basic mathematical morphology is also implemented in scipy.ndimage.morphology. The scipy.
[0, 0, 0, 0, 0, 0, 0], ndimage implementation works on arbitrary-dimensional arrays.
[0, 0, 0, 0, 0, 0, 0]], dtype=uint8)
>>> #Erosion removes objects smaller than the structure
>>> ski.morphology.binary_erosion(a, diamond(2)).astype(np.uint8)
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0], Example of filters comparison: image denoising
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0], >>> coins = ski.data.coins()
[0, 0, 0, 0, 0, 0, 0], >>> coins_zoom = coins[10:80, 300:370]
[0, 0, 0, 0, 0, 0, 0], >>> median_coins = ski.filters.median(
[0, 0, 0, 0, 0, 0, 0]], dtype=uint8) ... coins_zoom, disk(1)
... )
Dilation: maximum filter: >>> tv_coins = ski.restoration.denoise_tv_chambolle(
... coins_zoom, weight=0.1
>>> a = np.zeros((5, 5)) ... )
>>> a[2, 2] = 1 >>> gaussian_coins = ski.filters.gaussian(coins, sigma=2)
>>> a
array([[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 1., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.]])
>>> ski.morphology.binary_dilation(a, diamond(1)).astype(np.uint8)
array([[0, 0, 0, 0, 0],
[0, 0, 1, 0, 0],
[0, 1, 1, 1, 0],
[0, 0, 1, 0, 0],
[0, 0, 0, 0, 0]], dtype=uint8)

Opening: erosion + dilation: 17.6 Image segmentation

>>> a = np.zeros((5,5), dtype=int) Image segmentation is the attribution of different labels to different regions of the image, for example in
>>> a[1:4, 1:4] = 1; a[4, 4] = 1 order to extract the pixels of an object of interest.
>>> a
array([[0, 0, 0, 0, 0], 17.6.1 Binary segmentation: foreground + background
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0], Histogram-based method: Otsu thresholding
[0, 1, 1, 1, 0],
[0, 0, 0, 0, 1]])
 Tip
>>> ski.morphology.binary_opening(a, diamond(1)).astype(np.uint8)
array([[0, 0, 0, 0, 0], The Otsu method is a simple heuristic to find a threshold to separate the foreground from the
[0, 0, 1, 0, 0], background.
[0, 1, 1, 1, 0],
[0, 0, 1, 0, 0],
[0, 0, 0, 0, 0]], dtype=uint8)
Earlier scikit-image versions
Opening removes small objects and smoothes corners.

17.5. Image preprocessing / enhancement 580 17.6. Image segmentation 581

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

skimage.filters is called skimage.filter in earlier versions of scikit-image

camera = ski.data.camera()
val = ski.filters.threshold_otsu(camera)
mask = camera < val

ã See also

scipy.ndimage.find_objects() is useful to return slices on object in an image.

Labeling connected components of a discrete image 17.6.2 Marker based methods

If you have markers inside a set of regions, you can use these to segment the regions.
 Tip
Watershed segmentation
Once you have separated foreground objects, it is use to separate them from each other. For this, we
can assign a different integer labels to each one. The Watershed (skimage.segmentation.watershed()) is a region-growing approach that fills “basins”
in the image

Synthetic data: >>> # Generate an initial image with two overlapping circles
>>> x, y = np.indices((80, 80))
>>> n = 20 >>> x1, y1, x2, y2 = 28, 28, 44, 52
>>> l = 256 >>> r1, r2 = 16, 20
>>> im = np.zeros((l, l)) >>> mask_circle1 = (x - x1) ** 2 + (y - y1) ** 2 < r1 ** 2
>>> rng = np.random.default_rng() >>> mask_circle2 = (x - x2) ** 2 + (y - y2) ** 2 < r2 ** 2
>>> points = l * rng.random((2, n ** 2)) >>> image = np.logical_or(mask_circle1, mask_circle2)
>>> im[(points[0]).astype(int), (points[1]).astype(int)] = 1 >>> # Now we want to separate the two objects in image
>>> im = ski.filters.gaussian(im, sigma=l / (4. * n)) >>> # Generate the markers as local maxima of the distance
>>> blobs = im > im.mean() >>> # to the background
>>> import scipy as sp
Label all connected components: >>> distance = sp.ndimage.distance_transform_edt(image)
>>> peak_idx = ski.feature.peak_local_max(
>>> all_labels = ski.measure.label(blobs) ... distance, footprint=np.ones((3, 3)), labels=image
... )
Label only foreground connected components: >>> peak_mask = np.zeros_like(distance, dtype=bool)
>>> peak_mask[tuple(peak_idx.T)] = True
>>> blobs_labels = ski.measure.label(blobs, background=0)
>>> markers = ski.morphology.label(peak_mask)
>>> labels_ws = ski.segmentation.watershed(
... -distance, markers, mask=image
... )

Random walker segmentation

The random walker algorithm (skimage.segmentation.random_walker()) is similar to the Watershed,
but with a more “probabilistic” approach. It is based on the idea of the diffusion of labels in the image:

>>> # Transform markers image so that 0-valued pixels are to

>>> # be labelled, and -1-valued pixels represent background
(continues on next page)

17.6. Image segmentation 582 17.6. Image segmentation 583

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) 17.8 Data visualization and interaction

>>> markers[~image] = -1
>>> labels_rw = ski.segmentation.random_walker(image, markers) Meaningful visualizations are useful when testing a given processing pipeline.
Some image processing operations:

>>> coins = ski.data.coins()

>>> mask = coins > ski.filters.threshold_otsu(coins)
>>> clean_border = ski.segmentation.clear_border(mask)

Visualize binary result:

>>> plt.figure()
<Figure size ... with 0 Axes>
>>> plt.imshow(clean_border, cmap='gray')
<matplotlib.image.AxesImage object at 0x...>

Postprocessing label images Visualize contour

skimage provides several utility functions that can be used on label images (ie images where >>> plt.figure()
different discrete values identify different regions). Functions names are often self-explaining: <Figure size ... with 0 Axes>
skimage.segmentation.clear_border(), skimage.segmentation.relabel_from_one(), >>> plt.imshow(coins, cmap='gray')
skimage.morphology.remove_small_objects(), etc. <matplotlib.image.AxesImage object at 0x...>
>>> plt.contour(clean_border, [0.5])
<matplotlib.contour.QuadContourSet ...>
Exercise
Use skimage dedicated utility function:
• Load the coins image from the data submodule.
>>> coins_edges = ski.segmentation.mark_boundaries(
• Separate the coins from the background by testing several segmentation methods: Otsu thresh-
... coins, clean_border.astype(int)
olding, adaptive thresholding, and watershed or random walker segmentation.
... )
• If necessary, use a postprocessing function to improve the coins / background segmentation.

17.7 Measuring regions’ properties

Example: compute the size and perimeter of the two segmented regions:

>>> properties = ski.measure.regionprops(labels_rw)

>>> [float(prop.area) for prop in properties]
[770.0, 1168.0]
>>> [prop.perimeter for prop in properties]
[np.float64(100.91...), np.float64(126.81...)]

ã See also

for some properties, functions are available as well in scipy.ndimage.measurements with a different
API (a list is returned).
17.9 Feature extraction for computer vision
Geometric or textural descriptor can be extracted from images in order to
Exercise (continued)
• classify parts of the image (e.g. sky vs. buildings)
• Use the binary image of the coins and background from the previous exercise.
• match parts of different images (e.g. for object detection)
• Compute an image of labels for the different coins.
• and many other applications of Computer Vision
• Compute the size and eccentricity of all coins.
Example: detecting corners using Harris detector

17.7. Measuring regions’ properties 584 17.8. Data visualization and interaction 585
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

tform = ski.transform.AffineTransform(
scale=(1.3, 1.1), rotation=1, shear=0.7,
translation=(210, 50)
)
image = ski.transform.warp(
data.checkerboard(), tform.inverse, output_shape=(350, 350)
)

coords = ski.feature.corner_peaks(
ski.feature.corner_harris(image), min_distance=5
)
coords_subpix = ski.feature.corner_subpix(
image, coords, window_size=13
)

import numpy as np
import matplotlib.pyplot as plt

check = np.zeros((8, 8))

check[::2, 1::2] = 1
check[1::2, ::2] = 1
plt.matshow(check, cmap="gray")
plt.show()

Total running time of the script: (0 minutes 0.049 seconds)

(this example is taken from the plot_corner example in scikit-image)

Points of interest such as corners can then be used to match objects in different images, as described in 17.11.2 Displaying a simple image
the plot_matching example of scikit-image. Load and display an image

17.10 Full code examples

17.11 Examples for the scikit-image chapter
17.11.1 Creating an image
How to create an image with basic NumPy commands : np.zeros, slicing. . .
This examples show how to create a simple checkerboard.

17.10. Full code examples 586 17.11. Examples for the scikit-image chapter 587
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

import matplotlib.pyplot as plt

from skimage import data

camera = data.camera()
camera_multiply = 3 * camera
import matplotlib.pyplot as plt
from skimage import data plt.figure(figsize=(8, 4))
plt.subplot(121)
camera = data.camera() plt.imshow(camera, cmap="gray", interpolation="nearest")
plt.axis("off")
plt.subplot(122)
plt.figure(figsize=(4, 4)) plt.imshow(camera_multiply, cmap="gray", interpolation="nearest")
plt.imshow(camera, cmap="gray", interpolation="nearest") plt.axis("off")
plt.axis("off")
plt.tight_layout()
plt.tight_layout() plt.show()
plt.show()
Total running time of the script: (0 minutes 0.127 seconds)
Total running time of the script: (0 minutes 0.083 seconds)

17.11.4 Equalizing the histogram of an image

17.11.3 Integers can overflow Histogram equalizing makes images have a uniform histogram.
An illustration of overflow problem arising when working with integers

17.11. Examples for the scikit-image chapter 588 17.11. Examples for the scikit-image chapter 589
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

plt.figure(figsize=(12, 3))

plt.subplot(121)
plt.imshow(text, cmap="gray", interpolation="nearest")
plt.axis("off")
plt.subplot(122)
plt.imshow(hsobel_text, cmap="nipy_spectral", interpolation="nearest")
plt.axis("off")
plt.tight_layout()
plt.show()

Total running time of the script: (0 minutes 0.102 seconds)

from skimage import data, exposure 17.11.6 Segmentation contours

import matplotlib.pyplot as plt
Visualize segmentation contours on original grayscale image.
camera = data.camera()
camera_equalized = exposure.equalize_hist(camera)

plt.figure(figsize=(7, 3))

plt.subplot(121)
plt.imshow(camera, cmap="gray", interpolation="nearest")
plt.axis("off")
plt.subplot(122)
plt.imshow(camera_equalized, cmap="gray", interpolation="nearest")
plt.axis("off")
plt.tight_layout()
plt.show()

Total running time of the script: (0 minutes 0.096 seconds)

17.11.5 Computing horizontal gradients with the Sobel filter from skimage import data, segmentation
from skimage import filters
This example illustrates the use of the horizontal Sobel filter, to compute horizontal gradients. import matplotlib.pyplot as plt
import numpy as np

coins = data.coins()
mask = coins > filters.threshold_otsu(coins)
clean_border = segmentation.clear_border(mask).astype(int)

coins_edges = segmentation.mark_boundaries(coins, clean_border)

plt.figure(figsize=(8, 3.5))
plt.subplot(121)
plt.imshow(clean_border, cmap="gray")
from skimage import data plt.axis("off")
from skimage import filters plt.subplot(122)
import matplotlib.pyplot as plt plt.imshow(coins_edges)
plt.axis("off")
text = data.text()
hsobel_text = filters.sobel_h(text) (continues on next page)
(continues on next page)

17.11. Examples for the scikit-image chapter 590 17.11. Examples for the scikit-image chapter 591
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) 17.11.8 Affine transform

plt.tight_layout() Warping and affine transforms of images.
plt.show()

Total running time of the script: (0 minutes 0.103 seconds)

17.11.7 Otsu thresholding

This example illustrates automatic Otsu thresholding.

import matplotlib.pyplot as plt

from skimage import data
from skimage import filters
from skimage import exposure
import matplotlib.pyplot as plt
camera = data.camera()
from skimage import data
val = filters.threshold_otsu(camera)
from skimage.feature import corner_harris, corner_subpix, corner_peaks
from skimage.transform import warp, AffineTransform
hist, bins_center = exposure.histogram(camera)

plt.figure(figsize=(9, 4))
tform = AffineTransform(scale=(1.3, 1.1), rotation=1, shear=0.7, translation=(210,␣
plt.subplot(131)
˓→50))
plt.imshow(camera, cmap="gray", interpolation="nearest")
image = warp(data.checkerboard(), tform.inverse, output_shape=(350, 350))
plt.axis("off")
plt.subplot(132)
coords = corner_peaks(corner_harris(image), min_distance=5)
plt.imshow(camera < val, cmap="gray", interpolation="nearest")
coords_subpix = corner_subpix(image, coords, window_size=13)
plt.axis("off")
plt.subplot(133)
plt.gray()
plt.plot(bins_center, hist, lw=2)
plt.imshow(image, interpolation="nearest")
plt.axvline(val, color="k", ls="--")
plt.plot(coords_subpix[:, 1], coords_subpix[:, 0], "+r", markersize=15, mew=5)
plt.plot(coords[:, 1], coords[:, 0], ".b", markersize=7)
plt.tight_layout()
plt.axis("off")
plt.show()
plt.show()
Total running time of the script: (0 minutes 0.128 seconds)
Total running time of the script: (0 minutes 1.383 seconds)

17.11. Examples for the scikit-image chapter 592 17.11. Examples for the scikit-image chapter 593
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

17.11.9 Labelling connected components of an image 17.11.10 Various denoising filters

This example shows how to label connected components of a binary image, using the dedicated skim- This example compares several denoising filters available in scikit-image: a Gaussian filter, a median
age.measure.label function. filter, and total variation denoising.

import numpy as np
import matplotlib.pyplot as plt
from skimage import data
from skimage import filters
from skimage import restoration
from skimage import measure
from skimage import filters coins = data.coins()
import matplotlib.pyplot as plt gaussian_filter_coins = filters.gaussian(coins, sigma=2)
import numpy as np med_filter_coins = filters.median(coins, np.ones((3, 3)))
tv_filter_coins = restoration.denoise_tv_chambolle(coins, weight=0.1)
n = 12
l = 256 plt.figure(figsize=(16, 4))
rng = np.random.default_rng(27446968) plt.subplot(141)
im = np.zeros((l, l)) plt.imshow(coins[10:80, 300:370], cmap="gray", interpolation="nearest")
points = l * rng.random((2, n**2)) plt.axis("off")
im[(points[0]).astype(int), (points[1]).astype(int)] = 1 plt.title("Image")
im = filters.gaussian(im, sigma=l / (4.0 * n)) plt.subplot(142)
blobs = im > 0.7 * im.mean() plt.imshow(gaussian_filter_coins[10:80, 300:370], cmap="gray", interpolation="nearest
˓→")

all_labels = measure.label(blobs) plt.axis("off")

blobs_labels = measure.label(blobs, background=0) plt.title("Gaussian filter")
plt.subplot(143)
plt.figure(figsize=(9, 3.5)) plt.imshow(med_filter_coins[10:80, 300:370], cmap="gray", interpolation="nearest")
plt.subplot(131) plt.axis("off")
plt.imshow(blobs, cmap="gray") plt.title("Median filter")
plt.axis("off") plt.subplot(144)
plt.subplot(132) plt.imshow(tv_filter_coins[10:80, 300:370], cmap="gray", interpolation="nearest")
plt.imshow(all_labels, cmap="nipy_spectral") plt.axis("off")
plt.axis("off") plt.title("TV filter")
plt.subplot(133) plt.show()
plt.imshow(blobs_labels, cmap="nipy_spectral")
plt.axis("off") Total running time of the script: (0 minutes 0.153 seconds)

plt.tight_layout()
plt.show()
17.11.11 Watershed and random walker for segmentation
Total running time of the script: (0 minutes 0.076 seconds) This example compares two segmentation methods in order to separate two connected disks: the water-
shed algorithm, and the random walker algorithm.
Both segmentation methods require seeds, that are pixels belonging unambigusouly to a reagion. Here,
local maxima of the distance map to the background are used as seeds.

17.11. Examples for the scikit-image chapter 594 17.11. Examples for the scikit-image chapter 595
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

plt.tight_layout()
plt.show()

Total running time of the script: (0 minutes 0.168 seconds)

import numpy as np
from skimage.segmentation import watershed
from skimage.feature import peak_local_max
from skimage import measure
from skimage.segmentation import random_walker
import matplotlib.pyplot as plt
import scipy as sp

# Generate an initial image with two overlapping circles

x, y = np.indices((80, 80))
x1, y1, x2, y2 = 28, 28, 44, 52
r1, r2 = 16, 20
mask_circle1 = (x - x1) ** 2 + (y - y1) ** 2 < r1**2
mask_circle2 = (x - x2) ** 2 + (y - y2) ** 2 < r2**2
image = np.logical_or(mask_circle1, mask_circle2)
# Now we want to separate the two objects in image
# Generate the markers as local maxima of the distance
# to the background
distance = sp.ndimage.distance_transform_edt(image)
peak_idx = peak_local_max(distance, footprint=np.ones((3, 3)), labels=image)
peak_mask = np.zeros_like(distance, dtype=bool)
peak_mask[tuple(peak_idx.T)] = True
markers = measure.label(peak_mask)
labels_ws = watershed(-distance, markers, mask=image)

markers[~image] = -1
labels_rw = random_walker(image, markers)

plt.figure(figsize=(12, 3.5))
plt.subplot(141)
plt.imshow(image, cmap="gray", interpolation="nearest")
plt.axis("off")
plt.title("image")
plt.subplot(142)
plt.imshow(-distance, interpolation="nearest")
plt.axis("off")
plt.title("distance map")
plt.subplot(143)
plt.imshow(labels_ws, cmap="nipy_spectral", interpolation="nearest")
plt.axis("off")
plt.title("watershed segmentation")
plt.subplot(144)
plt.imshow(labels_rw, cmap="nipy_spectral", interpolation="nearest")
plt.axis("off")
plt.title("random walker segmentation")
(continues on next page)

17.11. Examples for the scikit-image chapter 596 17.11. Examples for the scikit-image chapter 597
 Scientific Python Lectures, Edition 2025.1rc0.dev0

• The documentation of scikit-learn is very complete and didactic.

Chapters contents

18
• Introduction: problem settings
• Basic principles of machine learning with scikit-learn
• Supervised Learning: Classification of Handwritten Digits
• Supervised Learning: Regression of Housing Data
CHAPTER
• Measuring prediction performance
• Unsupervised Learning: Dimensionality Reduction and Visualization
• Parameter selection, Validation, and Testing
• Examples for the scikit-learn chapter

scikit-learn: machine learning in Python 18.1 Introduction: problem settings

18.1.1 What is machine learning?

 Tip

Machine Learning is about building programs with tunable parameters that are adjusted auto-
matically so as to improve their behavior by adapting to previously seen data.
Authors: Gael Varoquaux
Machine Learning can be considered a subfield of Artificial Intelligence since those algorithms
can be seen as building blocks to make computers learn to behave more intelligently by somehow
generalizing rather that just storing and retrieving data items like a database system would do.

Prerequisites
• numpy
• scipy
• matplotlib (optional)
• ipython (the enhancements come handy)

Acknowledgements

This chapter is adapted from a tutorial given by Gaël Varoquaux, Jake Vanderplas, Olivier Grisel.

ã See also Fig. 1: A classification problem

Data science in Python
We’ll take a look at two very simple machine learning tasks here. The first is a classification task:
• The Statistics in Python chapter may also be of interest for readers looking into machine
the figure shows a collection of two-dimensional data, colored according to two different class labels. A
learning.
classification algorithm may be used to draw a dividing boundary between the two clusters of points:

598 18.1. Introduction: problem settings 599

 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

By drawing this separating line, we have learned a model which can generalize to new data: if you were A Simple Example: the Iris Dataset
to drop another point onto the plane which is unlabeled, this algorithm could now predict whether it’s
The application problem
a blue or a red point.
As an example of a simple dataset, let us a look at the iris data stored by scikit-learn. Suppose we want
to recognize species of irises. The data consists of measurements of three different species of irises:

Setosa Iris Versicolor Iris Virginica Iris

Fig. 2: A regression problem Quick Question:

If we want to design an algorithm to recognize iris species, what might the

data be?
Remember: we need a 2D array of size [n_samples x n_features].
• What would the n_samples refer to?
• What might the n_features refer to?
The next simple task we’ll look at is a regression task: a simple best-fit line to a set of data.
Remember that there must be a fixed number of features for each sample, and feature number i must
Again, this is an example of fitting a model to data, but our focus here is that the model can make be a similar kind of quantity for each sample.
generalizations about new data. The model has been learned from the training data, and can be used
to predict the result of test data: here, we might be given an x-value, and the model would allow us to Loading the Iris Data with Scikit-learn
predict the y value.
Scikit-learn has a very straightforward set of data on these iris species. The data consist of the following:
18.1.2 Data in scikit-learn • Features in the Iris dataset:
The data matrix – sepal length (cm)
Machine learning algorithms implemented in scikit-learn expect data to be stored in a two-dimensional – sepal width (cm)
array or matrix. The arrays can be either numpy arrays, or in some cases scipy.sparse matrices. The
– petal length (cm)
size of the array is expected to be [n_samples, n_features]
– petal width (cm)
• n_samples: The number of samples: each sample is an item to process (e.g. classify). A sample
can be a document, a picture, a sound, a video, an astronomical object, a row in database or CSV • Target classes to predict:
file, or whatever you can describe with a fixed set of quantitative traits.
– Setosa
• n_features: The number of features or distinct traits that can be used to describe each item in
– Versicolour
a quantitative manner. Features are generally real-valued, but may be boolean or discrete-valued
in some cases. – Virginica
scikit-learn embeds a copy of the iris CSV file along with a function to load it into NumPy arrays:
 Tip
>>> from sklearn.datasets import load_iris
The number of features must be fixed in advance. However it can be very high dimensional (e.g. >>> iris = load_iris()
millions of features) with most of them being zeros for a given sample. This is a case where scipy.
sparse matrices can be useful, in that they are much more memory-efficient than NumPy arrays.

18.1. Introduction: problem settings 600 18.1. Introduction: problem settings 601
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

ò Note
Exercise:
Import sklearn Note that scikit-learn is imported as sklearn
Can you choose 2 features to find a plot where it is easier to separate the different classes of irises?
Hint: click on the figure above to see the code that generates it, and modify this code.
The features of each sample flower are stored in the data attribute of the dataset:

>>> print(iris.data.shape)
(150, 4) 18.2 Basic principles of machine learning with scikit-learn
>>> n_samples, n_features = iris.data.shape
>>> print(n_samples) 18.2.1 Introducing the scikit-learn estimator object
150
>>> print(n_features) Every algorithm is exposed in scikit-learn via an ‘’Estimator” object. For instance a linear regression is:
4 sklearn.linear_model.LinearRegression
>>> print(iris.data[0])
>>> from sklearn.linear_model import LinearRegression
[5.1 3.5 1.4 0.2]
Estimator parameters: All the parameters of an estimator can be set when it is instantiated:
The information about the class of each sample is stored in the target attribute of the dataset:
>>> model = LinearRegression(n_jobs=1)
>>> print(iris.target.shape)
>>> print(model)
(150,)
LinearRegression(n_jobs=1)
>>> print(iris.target)
[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 Fitting on data
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 Let’s create some simple data with numpy:
2 2]
>>> import numpy as np
>>> x = np.array([0, 1, 2])
The names of the classes are stored in the last attribute, namely target_names:
>>> y = np.array([0, 1, 2])
>>> print(iris.target_names)
['setosa' 'versicolor' 'virginica'] >>> X = x[:, np.newaxis] # The input data for sklearn is 2D: (samples == 3 x features␣
˓→== 1)

This data is four-dimensional, but we can visualize two of the dimensions at a time using a scatter plot: >>> X
array([[0],
[1],
[2]])

>>> model.fit(X, y)
LinearRegression(n_jobs=1)

Estimated parameters: When data is fitted with an estimator, parameters are estimated from the data
at hand. All the estimated parameters are attributes of the estimator object ending by an underscore:

>>> model.coef_
array([1.])

18.2.2 Supervised Learning: Classification and regression

In Supervised Learning, we have a dataset consisting of both features and labels. The task is to
construct an estimator which is able to predict the label of an object given the set of features. A
relatively simple example is predicting the species of iris given a set of measurements of its flower. This
is a relatively simple task. Some more complicated examples are:
• given a multicolor image of an object through a telescope, determine whether that object is a star,
a quasar, or a galaxy.
• given a photograph of a person, identify the person in the photo.

18.1. Introduction: problem settings 602 18.2. Basic principles of machine learning with scikit-learn 603
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• given a list of movies a person has watched and their personal rating of the movie, recommend a
list of movies they would like (So-called recommender systems: a famous example is the Netflix
Prize).

 Tip

What these tasks have in common is that there is one or more unknown quantities associated with
the object which needs to be determined from other observed quantities.

Supervised learning is further broken down into two categories, classification and regression. In
classification, the label is discrete, while in regression, the label is continuous. For example, in astronomy,
the task of determining whether an object is a star, a galaxy, or a quasar is a classification problem: the
label is from three distinct categories. On the other hand, we might wish to estimate the age of an object
based on such observations: this would be a regression problem, because the label (age) is a continuous
quantity.
Classification: K nearest neighbors (kNN) is one of the simplest learning strategies: given a new,
unknown observation, look up in your reference database which ones have the closest features and assign
the predominant class. Let’s try it out on our iris classification problem:

from sklearn import neighbors, datasets

iris = datasets.load_iris()
X, y = iris.data, iris.target
knn = neighbors.KNeighborsClassifier(n_neighbors=1)
knn.fit(X, y)
# What kind of iris has 3cm x 5cm sepal and 4cm x 2cm petal?
print(iris.target_names[knn.predict([[3, 5, 4, 2]])])

Regression: The simplest possible regression setting is the linear regression one:

from sklearn.linear_model import LinearRegression

Fig. 3: A plot of the sepal space and the prediction of the KNN
# x from 0 to 30
rng = np.random.default_rng()
x = 30 * rng.random((20, 1))

# y = a*x + b with noise

y = 0.5 * x + 1.0 + rng.normal(size=x.shape)

# create a linear regression model

model = LinearRegression()
model.fit(x, y)

# predict y from the data

x_new = np.linspace(0, 30, 100)
y_new = model.predict(x_new[:, np.newaxis])

18.2.3 A recap on Scikit-learn’s estimator interface

Scikit-learn strives to have a uniform interface across all methods, and we’ll see examples of these below.
Given a scikit-learn estimator object named model, the following methods are available:
In all Estimators
• model.fit() : fit training data. For supervised learning applications, this accepts
two arguments: the data X and the labels y (e.g. model.fit(X, y)). For unsu- Fig. 4: A plot of a simple linear regression.
pervised learning applications, this accepts only a single argument, the data X (e.g.
model.fit(X)).

18.2. Basic principles of machine learning with scikit-learn 604 18.2. Basic principles of machine learning with scikit-learn 605
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

In supervised estimators And now, let’s fit a 4th order and a 9th order polynomial to the data.
• model.predict() : given a trained model, predict the label of a new set of
data. This method accepts one argument, the new data X_new (e.g. model.
predict(X_new)), and returns the learned label for each object in the array.
• model.predict_proba() : For classification problems, some estimators also pro-
vide this method, which returns the probability that a new observation has each
categorical label. In this case, the label with the highest probability is returned by
model.predict().
• model.score() : for classification or regression problems, most (all?) estimators
implement a score method. Scores are between 0 and 1, with a larger score indicating
a better fit.
In unsupervised estimators
• model.transform() : given an unsupervised model, transform new data into the
new basis. This also accepts one argument X_new, and returns the new representa-
tion of the data based on the unsupervised model.
• model.fit_transform() : some estimators implement this method, which more
efficiently performs a fit and a transform on the same input data.

18.2.4 Regularization: what it is and why it is necessary

Preferring simpler models
Train errors Suppose you are using a 1-nearest neighbor estimator. How many errors do you expect With your naked eyes, which model do you prefer, the 4th order one, or the 9th order one?
on your train set?
Let’s look at the ground truth:
• Train set error is not a good measurement of prediction performance. You need to leave out a test
set.
• In general, we should accept errors on the train set.
An example of regularization The core idea behind regularization is that we are going to prefer
models that are simpler, for a certain definition of ‘’simpler”, even if they lead to more errors on the
train set.
As an example, let’s generate with a 9th order polynomial, with noise:

 Tip

Regularization is ubiquitous in machine learning. Most scikit-learn estimators have a parameter to

tune the amount of regularization. For instance, with k-NN, it is ‘k’, the number of nearest neighbors
used to make the decision. k=1 amounts to no regularization: 0 error on the training set, whereas
large k will push toward smoother decision boundaries in the feature space.

18.2. Basic principles of machine learning with scikit-learn 606 18.2. Basic principles of machine learning with scikit-learn 607
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Simple versus complex models for classification

A linear separation A non-linear separation

 Tip

For classification models, the decision boundary, that separates the class expresses the complexity
of the model. For instance, a linear model, that makes a decision based on a linear combination of
features, is more complex than a non-linear one.

18.3 Supervised Learning: Classification of Handwritten Digits

18.3.1 The nature of the data

Code and notebook

Python code and Jupyter notebook for this section are found here

In this section we’ll apply scikit-learn to the classification of handwritten digits. This will go a bit beyond
the iris classification we saw before: we’ll discuss some of the metrics which can be used in evaluating
the effectiveness of a classification model. Let us visualize the data and remind us what we’re looking at (click on the figure for the full code):

>>> from sklearn.datasets import load_digits # plot the digits: each image is 8x8 pixels
>>> digits = load_digits() for i in range(64):
ax = fig.add_subplot(8, 8, i + 1, xticks=[], yticks=[])
ax.imshow(digits.images[i], cmap=plt.cm.binary, interpolation='nearest')

18.3.2 Visualizing the Data on its principal components

A good first-step for many problems is to visualize the data using a Dimensionality Reduction technique.
We’ll start with the most straightforward one, Principal Component Analysis (PCA).
PCA seeks orthogonal linear combinations of the features which show the greatest variance, and as such,
can help give you a good idea of the structure of the data set.

>>> from sklearn.decomposition import PCA

>>> pca = PCA(n_components=2)
>>> proj = pca.fit_transform(digits.data)
>>> plt.scatter(proj[:, 0], proj[:, 1], c=digits.target)
<matplotlib.collections.PathCollection object at ...>
>>> plt.colorbar()
(continues on next page)

18.3. Supervised Learning: Classification of Handwritten Digits 608 18.3. Supervised Learning: Classification of Handwritten Digits 609
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> from sklearn.naive_bayes import GaussianNB
<matplotlib.colorbar.Colorbar object at ...> >>> from sklearn.model_selection import train_test_split

>>> # split the data into training and validation sets

>>> X_train, X_test, y_train, y_test = train_test_split(
... digits.data, digits.target, random_state=42)

>>> # train the model

>>> clf = GaussianNB()
>>> clf.fit(X_train, y_train)
GaussianNB()

>>> # use the model to predict the labels of the test data
>>> predicted = clf.predict(X_test)
>>> expected = y_test
>>> print(predicted)
[6 9 3 7 2 2 5 8 5 2 1 1 7 0 4 8 3 7 8 8 4 3 9 7 5 6 3 5 6 3...]
>>> print(expected)
[6 9 3 7 2 1 5 2 5 2 1 9 4 0 4 2 3 7 8 8 4 3 9 7 5 6 3 5 6 3...]

As above, we plot the digits with the predicted labels to get an idea of how well the classification is
working.

Question

Given these projections of the data, which numbers do you think a classifier might have trouble
distinguishing?

18.3.3 Gaussian Naive Bayes Classification

For most classification problems, it’s nice to have a simple, fast method to provide a quick baseline
classification. If the simple and fast method is sufficient, then we don’t have to waste CPU cycles on
more complex models. If not, we can use the results of the simple method to give us clues about our
data.
One good method to keep in mind is Gaussian Naive Bayes (sklearn.naive_bayes.GaussianNB).

Old scikit-learn versions

train_test_split() is imported from sklearn.cross_validation

 Tip

Gaussian Naive Bayes fits a Gaussian distribution to each training label independently on each
feature, and uses this to quickly give a rough classification. It is generally not sufficiently accurate
for real-world data, but can perform surprisingly well, for instance on text data.

18.3. Supervised Learning: Classification of Handwritten Digits 610 18.3. Supervised Learning: Classification of Handwritten Digits 611
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

One of the most useful metrics is the classification_report, which combines several measures and
prints a table with the results:

>>> from sklearn import metrics

>>> print(metrics.classification_report(expected, predicted))
precision recall f1-score support

0 1.00 0.95 0.98 43

1 0.85 0.78 0.82 37
2 0.85 0.61 0.71 38
3 0.97 0.83 0.89 46
4 0.98 0.84 0.90 55
5 0.90 0.95 0.93 59
6 0.90 0.96 0.92 45
7 0.71 0.98 0.82 41
8 0.60 0.89 0.72 38
9 0.90 0.73 0.80 48

accuracy 0.86 450

macro avg 0.87 0.85 0.85 450
weighted avg 0.88 0.86 0.86 450

Another enlightening metric for this sort of multi-label classification is a confusion matrix: it helps us
visualize which labels are being interchanged in the classification errors:

>>> print(metrics.confusion_matrix(expected, predicted))

[[41 0 0 0 0 1 0 1 0 0]
[ 0 29 2 0 0 0 0 0 4 2]
[ 0 2 23 0 0 0 1 0 12 0]
[ 0 0 1 38 0 1 0 0 5 1]
[ 0 0 0 0 46 0 2 7 0 0]
[ 0 0 0 0 0 56 1 1 0 1]
[ 0 0 0 0 1 1 43 0 0 0]
[ 0 0 0 0 0 1 0 40 0 0]
[ 0 2 0 0 0 0 0 2 34 0]
[ 0 1 1 1 0 2 1 5 2 35]]

We see here that in particular, the numbers 1, 2, 3, and 9 are often being labeled 8.
Question

Why did we split the data into training and validation sets? 18.4 Supervised Learning: Regression of Housing Data
Here we’ll do a short example of a regression problem: learning a continuous value from a set of features.
18.3.4 Quantitative Measurement of Performance
18.4.1 A quick look at the data
We’d like to measure the performance of our estimator without having to resort to plotting examples. A
simple method might be to simply compare the number of matches:
Code and notebook
>>> matches = (predicted == expected)
>>> print(matches.sum()) Python code and Jupyter notebook for this section are found here
385
>>> print(len(matches)) We’ll use the California house prices set, available in scikit-learn. This records measurements of 8
450 attributes of housing markets in California, as well as the median price. The question is: can you predict
>>> matches.sum() / float(len(matches)) the price of a new market given its attributes?:
np.float64(0.8555...)
>>> from sklearn.datasets import fetch_california_housing
We see that more than 80% of the 450 predictions match the input. But there are other more sophisticated >>> data = fetch_california_housing(as_frame=True)
metrics that can be used to judge the performance of a classifier: several are available in the sklearn. >>> print(data.data.shape)
metrics submodule. (continues on next page)

18.3. Supervised Learning: Classification of Handwritten Digits 612 18.4. Supervised Learning: Regression of Housing Data 613
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> plt.hist(data.target)
(20640, 8) (array([...
>>> print(data.target.shape)
(20640,)

We can see that there are just over 20000 data points.
The DESCR variable has a long description of the dataset:

>>> print(data.DESCR)
.. _california_housing_dataset:

California Housing dataset

--------------------------

**Data Set Characteristics:**

:Number of Instances: 20640 Let’s have a quick look to see if some features are more relevant than others for our problem:
:Number of Attributes: 8 numeric, predictive attributes and the target >>> for index, feature_name in enumerate(data.feature_names):
... plt.figure()
:Attribute Information: ... plt.scatter(data.data[feature_name], data.target)
- MedInc median income in block group <Figure size...
- HouseAge median house age in block group
- AveRooms average number of rooms per household
- AveBedrms average number of bedrooms per household
- Population block group population
- AveOccup average number of household members
- Latitude block group latitude
- Longitude block group longitude

:Missing Attribute Values: None

This dataset was obtained from the StatLib repository.

https://www.dcc.fc.up.pt/~ltorgo/Regression/cal_housing.html

The target variable is the median house value for California districts,
expressed in hundreds of thousands of dollars ($100,000).

This dataset was derived from the 1990 U.S. census, using one row per census
block group. A block group is the smallest geographical unit for which the U.S.
Census Bureau publishes sample data (a block group typically has a population
of 600 to 3,000 people).

A household is a group of people residing within a home. Since the average

number of rooms and bedrooms in this dataset are provided per household, these
columns may take surprisingly large values for block groups with few households
and many empty houses, such as vacation resorts.

It can be downloaded/loaded using the

:func:`sklearn.datasets.fetch_california_housing` function.

.. rubric:: References

- Pace, R. Kelley and Ronald Barry, Sparse Spatial Autoregressions,

Statistics and Probability Letters, 33 (1997) 291-297

It often helps to quickly visualize pieces of the data using histograms, scatter plots, or other plot types.
With matplotlib, let us show a histogram of the target values: the median price in each neighborhood:

18.4. Supervised Learning: Regression of Housing Data 614 18.4. Supervised Learning: Regression of Housing Data 615
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

18.4.2 Predicting Home Prices: a Simple Linear Regression

Now we’ll use scikit-learn to perform a simple linear regression on the housing data. There are many
possibilities of regressors to use. A particularly simple one is LinearRegression: this is basically a
wrapper around an ordinary least squares calculation.
>>> from sklearn.model_selection import train_test_split
>>> X_train, X_test, y_train, y_test = train_test_split(data.data, data.target)
>>> from sklearn.linear_model import LinearRegression
>>> clf = LinearRegression()
>>> clf.fit(X_train, y_train)
LinearRegression()
>>> predicted = clf.predict(X_test)
>>> expected = y_test
>>> print("RMS: %s " % np.sqrt(np.mean((predicted - expected) ** 2)))
RMS: 0.7...

We can plot the error: expected as a function of predicted:

>>> plt.scatter(expected, predicted)
<matplotlib.collections.PathCollection object at ...>

 Tip

The prediction at least correlates with the true price, though there are clearly some biases. We could
imagine evaluating the performance of the regressor by, say, computing the RMS residuals between
the true and predicted price. There are some subtleties in this, however, which we’ll cover in a later
section.

This is a manual version of a technique called feature selection.

Exercise: Gradient Boosting Tree Regression

There are many other types of regressors available in scikit-learn: we’ll try a more powerful one here.
 Tip
Use the GradientBoostingRegressor class to fit the housing data.
Sometimes, in Machine Learning it is useful to use feature selection to decide which features are the
most useful for a particular problem. Automated methods exist which quantify this sort of exercise hint You can copy and paste some of the above code, replacing LinearRegression with
of choosing the most informative features. GradientBoostingRegressor:
from sklearn.ensemble import GradientBoostingRegressor
# Instantiate the model, fit the results, and scatter in vs. out

18.4. Supervised Learning: Regression of Housing Data 616 18.4. Supervised Learning: Regression of Housing Data 617
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Solution The solution is found in the code of this chapter

18.5 Measuring prediction performance

18.5.1 A quick test on the K-neighbors classifier
Here we’ll continue to look at the digits data, but we’ll switch to the K-Neighbors classifier. The K-
neighbors classifier is an instance-based classifier. The K-neighbors classifier predicts the label of an
unknown point based on the labels of the K nearest points in the parameter space.

>>> # Get the data

>>> from sklearn.datasets import load_digits
>>> digits = load_digits()
>>> X = digits.data
>>> y = digits.target

>>> # Instantiate and train the classifier

>>> from sklearn.neighbors import KNeighborsClassifier Here again the predictions are seemingly perfect as the model was able to perfectly memorize the training
>>> clf = KNeighborsClassifier(n_neighbors=1) set.
>>> clf.fit(X, y)
KNeighborsClassifier(...)
. Warning
>>> # Check the results using metrics
Performance on test set
>>> from sklearn import metrics
>>> y_pred = clf.predict(X) Performance on test set does not measure overfit (as described above)

>>> print(metrics.confusion_matrix(y_pred, y))

[[178 0 0 0 0 0 0 0 0 0] 18.5.2 A correct approach: Using a validation set
[ 0 182 0 0 0 0 0 0 0 0]
[ 0 0 177 0 0 0 0 0 0 0] Learning the parameters of a prediction function and testing it on the same data is a methodological
[ 0 0 0 183 0 0 0 0 0 0] mistake: a model that would just repeat the labels of the samples that it has just seen would have a
[ 0 0 0 0 181 0 0 0 0 0] perfect score but would fail to predict anything useful on yet-unseen data.
[ 0 0 0 0 0 182 0 0 0 0] To avoid over-fitting, we have to define two different sets:
[ 0 0 0 0 0 0 181 0 0 0]
[ 0 0 0 0 0 0 0 179 0 0] • a training set X_train, y_train which is used for learning the parameters of a predictive model
[ 0 0 0 0 0 0 0 0 174 0] • a testing set X_test, y_test which is used for evaluating the fitted predictive model
[ 0 0 0 0 0 0 0 0 0 180]]
In scikit-learn such a random split can be quickly computed with the train_test_split() function:
Apparently, we’ve found a perfect classifier! But this is misleading for the reasons we saw before: the >>> from sklearn import model_selection
classifier essentially “memorizes” all the samples it has already seen. To really test how well this algorithm >>> X = digits.data
does, we need to try some samples it hasn’t yet seen. >>> y = digits.target
This problem also occurs with regression models. In the following we fit an other instance-based model
named “decision tree” to the California Housing price dataset we introduced previously: >>> X_train, X_test, y_train, y_test = model_selection.train_test_split(X, y,
... test_size=0.25, random_state=0)
>>> from sklearn.datasets import fetch_california_housing
>>> from sklearn.tree import DecisionTreeRegressor >>> print("%r , %r , %r " % (X.shape, X_train.shape, X_test.shape))
(1797, 64), (1347, 64), (450, 64)
>>> data = fetch_california_housing(as_frame=True)
>>> clf = DecisionTreeRegressor().fit(data.data, data.target) Now we train on the training data, and test on the testing data:
>>> predicted = clf.predict(data.data)
>>> expected = data.target >>> clf = KNeighborsClassifier(n_neighbors=1).fit(X_train, y_train)
>>> y_pred = clf.predict(X_test)
>>> plt.scatter(expected, predicted)
<matplotlib.collections.PathCollection object at ...> >>> print(metrics.confusion_matrix(y_test, y_pred))
>>> plt.plot([0, 50], [0, 50], '--k') [[37 0 0 0 0 0 0 0 0 0]
[<matplotlib.lines.Line2D object at ...] [ 0 43 0 0 0 0 0 0 0 0]
(continues on next page)

18.5. Measuring prediction performance 618 18.5. Measuring prediction performance 619
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> from sklearn.naive_bayes import GaussianNB
[ 0 0 43 1 0 0 0 0 0 0] >>> from sklearn.neighbors import KNeighborsClassifier
[ 0 0 0 45 0 0 0 0 0 0] >>> from sklearn.svm import LinearSVC
[ 0 0 0 0 38 0 0 0 0 0]
[ 0 0 0 0 0 47 0 0 0 1] >>> X = digits.data
[ 0 0 0 0 0 0 52 0 0 0] >>> y = digits.target
[ 0 0 0 0 0 0 0 48 0 0] >>> X_train, X_test, y_train, y_test = model_selection.train_test_split(X, y,
[ 0 0 0 0 0 0 0 0 48 0] ... test_size=0.25, random_state=0)
[ 0 0 0 1 0 1 0 0 0 45]]
>>> print(metrics.classification_report(y_test, y_pred)) >>> for Model in [GaussianNB(), KNeighborsClassifier(), LinearSVC(dual=False)]:
precision recall f1-score support ... clf = Model.fit(X_train, y_train)
... y_pred = clf.predict(X_test)
0 1.00 1.00 1.00 37 ... print('%s : %s ' %
1 1.00 1.00 1.00 43 ... (Model.__class__.__name__, metrics.f1_score(y_test, y_pred,␣
2 1.00 0.98 0.99 44 ˓→average="macro")))
3 0.96 1.00 0.98 45 GaussianNB: 0.8...
4 1.00 1.00 1.00 38 KNeighborsClassifier: 0.9...
5 0.98 0.98 0.98 48 LinearSVC: 0.9...
6 1.00 1.00 1.00 52
7 1.00 1.00 1.00 48 • For each classifier, which value for the hyperparameters gives the best results for the digits data?
8 1.00 1.00 1.00 48 For LinearSVC, use loss='l2' and loss='l1'. For KNeighborsClassifier we use n_neighbors
9 0.98 0.96 0.97 47 between 1 and 10. Note that GaussianNB does not have any adjustable hyperparameters.
accuracy 0.99 450 LinearSVC(loss='l1'): 0.930570687535
macro avg 0.99 0.99 0.99 450 LinearSVC(loss='l2'): 0.933068826918
weighted avg 0.99 0.99 0.99 450 -------------------
KNeighbors(n_neighbors=1): 0.991367521884
The averaged f1-score is often used as a convenient measure of the overall performance of an algorithm. KNeighbors(n_neighbors=2): 0.984844206884
It appears in the bottom row of the classification report; it can also be accessed directly: KNeighbors(n_neighbors=3): 0.986775344954
KNeighbors(n_neighbors=4): 0.980371905382
>>> metrics.f1_score(y_test, y_pred, average="macro") KNeighbors(n_neighbors=5): 0.980456280495
0.991367... KNeighbors(n_neighbors=6): 0.975792419414
KNeighbors(n_neighbors=7): 0.978064579214
The over-fitting we saw previously can be quantified by computing the f1-score on the training data KNeighbors(n_neighbors=8): 0.978064579214
itself: KNeighbors(n_neighbors=9): 0.978064579214
KNeighbors(n_neighbors=10): 0.975555089773
>>> metrics.f1_score(y_train, clf.predict(X_train), average="macro")
1.0
Solution: code source

ò Note 18.5.4 Cross-validation

Cross-validation consists in repeatedly splitting the data in pairs of train and test sets, called ‘folds’.
Regression metrics In the case of regression models, we need to use different metrics, such as
Scikit-learn comes with a function to automatically compute score on all these folds. Here we do KFold
explained variance.
with k=5.

>>> clf = KNeighborsClassifier()

18.5.3 Model Selection via Validation >>> from sklearn.model_selection import cross_val_score
>>> cross_val_score(clf, X, y, cv=5)
array([0.947..., 0.955..., 0.966..., 0.980..., 0.963... ])
 Tip

We have applied Gaussian Naives, support vectors machines, and K-nearest neighbors classifiers to We can use different splitting strategies, such as random splitting:
the digits dataset. Now that we have these validation tools in place, we can ask quantitatively which
>>> from sklearn.model_selection import ShuffleSplit
of the three estimators works best for this dataset.
>>> cv = ShuffleSplit(n_splits=5)
>>> cross_val_score(clf, X, y, cv=cv)
• With the default hyper-parameters for each estimator, which gives the best f1 score on the valida- array([...])
tion set? Recall that hyperparameters are the parameters set when you instantiate the classifier:
for example, the n_neighbors in clf = KNeighborsClassifier(n_neighbors=1)

18.5. Measuring prediction performance 620 18.5. Measuring prediction performance 621
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

 Tip Question

There exists many different cross-validation strategies in scikit-learn. They are often useful to take Can we trust our results to be actually useful?
in account non iid datasets.
Automatically Performing Grid Search
18.5.5 Hyperparameter optimization with cross-validation sklearn.grid_search.GridSearchCV is constructed with an estimator, as well as a dictionary of pa-
Consider regularized linear models, such as Ridge Regression, which uses l2 regularization, and Lasso rameter values to be searched. We can find the optimal parameters this way:
Regression, which uses l1 regularization. Choosing their regularization parameter is important. >>> from sklearn.model_selection import GridSearchCV
Let us set these parameters on the Diabetes dataset, a simple regression problem. The diabetes data >>> for Model in [Ridge, Lasso]:
consists of 10 physiological variables (age, sex, weight, blood pressure) measure on 442 patients, and an ... gscv = GridSearchCV(Model(), dict(alpha=alphas), cv=3).fit(X, y)
indication of disease progression after one year: ... print('%s : %s ' % (Model.__name__, gscv.best_params_))
Ridge: {'alpha': np.float64(0.06210169418915616)}
>>> from sklearn.datasets import load_diabetes Lasso: {'alpha': np.float64(0.01268961003167922)}
>>> data = load_diabetes()
>>> X, y = data.data, data.target
>>> print(X.shape) Built-in Hyperparameter Search
(442, 10)
For some models within scikit-learn, cross-validation can be performed more efficiently on large datasets.
In this case, a cross-validated version of the particular model is included. The cross-validated versions
With the default hyper-parameters: we compute the cross-validation score: of Ridge and Lasso are RidgeCV and LassoCV, respectively. Parameter search on these estimators can
>>> from sklearn.linear_model import Ridge, Lasso be performed as follows:

>>> from sklearn.linear_model import RidgeCV, LassoCV

>>> for Model in [Ridge, Lasso]: >>> for Model in [RidgeCV, LassoCV]:
... model = Model() ... model = Model(alphas=alphas, cv=3).fit(X, y)
... print('%s : %s ' % (Model.__name__, cross_val_score(model, X, y).mean())) ... print('%s : %s ' % (Model.__name__, model.alpha_))
Ridge: 0.4... RidgeCV: 0.0621016941892
Lasso: 0.3... LassoCV: 0.0126896100317

Basic Hyperparameter Optimization We see that the results match those returned by GridSearchCV

We compute the cross-validation score as a function of alpha, the strength of the regularization for Lasso Nested cross-validation
and Ridge. We choose 20 values of alpha between 0.0001 and 1:
How do we measure the performance of these estimators? We have used data to set the hyperparameters,
>>> alphas = np.logspace(-3, -1, 30) so we need to test on actually new data. We can do this by running cross_val_score() on our CV
objects. Here there are 2 cross-validation loops going on, this is called ‘nested cross validation’:
>>> for Model in [Lasso, Ridge]:
... scores = [cross_val_score(Model(alpha), X, y, cv=3).mean() for Model in [RidgeCV, LassoCV]:
... for alpha in alphas] scores = cross_val_score(Model(alphas=alphas, cv=3), X, y, cv=3)
... plt.plot(alphas, scores, label=Model.__name__) print(Model.__name__, np.mean(scores))
[<matplotlib.lines.Line2D object at ...

ò Note

Note that these results do not match the best results of our curves above, and LassoCV seems to
under-perform RidgeCV. The reason is that setting the hyper-parameter is harder for Lasso, thus the
estimation error on this hyper-parameter is larger.

18.6 Unsupervised Learning: Dimensionality Reduction and Visual-

ization
Unsupervised learning is applied on X without y: data without labels. A typical use case is to find
hidden structure in the data.

18.5. Measuring prediction performance 622 18.6. Unsupervised Learning: Dimensionality Reduction and Visualization 623
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

18.6.1 Dimensionality Reduction: PCA

Dimensionality reduction derives a set of new artificial features smaller than the original feature set.
Here we’ll use Principal Component Analysis (PCA), a dimensionality reduction that strives to retain
most of the variance of the original data. We’ll use sklearn.decomposition.PCA on the iris dataset:

>>> X = iris.data
>>> y = iris.target

 Tip

PCA computes linear combinations of the original features using a truncated Singular Value Decom-
position of the matrix X, to project the data onto a base of the top singular vectors.

>>> from sklearn.decomposition import PCA

>>> pca = PCA(n_components=2, whiten=True)
>>> pca.fit(X)
PCA(n_components=2, whiten=True)

Once fitted, PCA exposes the singular vectors in the components_ attribute:

>>> pca.components_
array([[ 0.3..., -0.08..., 0.85..., 0.3...],  Tip
[ 0.6..., 0.7..., -0.1..., -0.07...]])
Note that this projection was determined without any information about the labels (represented by
Other attributes are available as well: the colors): this is the sense in which the learning is unsupervised. Nevertheless, we see that the
projection gives us insight into the distribution of the different flowers in parameter space: notably,
>>> pca.explained_variance_ratio_ iris setosa is much more distinct than the other two species.
array([0.92..., 0.053...])

Let us project the iris dataset along those first two dimensions:: 18.6.2 Visualization with a non-linear embedding: tSNE
>>> X_pca = pca.transform(X) For visualization, more complex embeddings can be useful (for statistical analysis, they are harder to
>>> X_pca.shape control). sklearn.manifold.TSNE is such a powerful manifold learning method. We apply it to the digits
(150, 2) dataset, as the digits are vectors of dimension 8*8 = 64. Embedding them in 2D enables visualization:

>>> # Take the first 500 data points: it's hard to see 1500 points
PCA normalizes and whitens the data, which means that the data is now centered on both components >>> X = digits.data[:500]
with unit variance: >>> y = digits.target[:500]
>>> X_pca.mean(axis=0)
array([...e-15, ...e-15]) >>> # Fit and transform with a TSNE
>>> X_pca.std(axis=0, ddof=1) >>> from sklearn.manifold import TSNE
array([1., 1.]) >>> tsne = TSNE(n_components=2, learning_rate='auto', init='random', random_state=0)
>>> X_2d = tsne.fit_transform(X)
Furthermore, the samples components do no longer carry any linear correlation:
>>> # Visualize the data
>>> np.corrcoef(X_pca.T) >>> plt.scatter(X_2d[:, 0], X_2d[:, 1], c=y)
array([[1.00000000e+00, 0.0], <matplotlib.collections.PathCollection object at ...>
[0.0, 1.00000000e+00]])

With a number of retained components 2 or 3, PCA is useful to visualize the dataset:

>>> target_ids = range(len(iris.target_names))

>>> for i, c, label in zip(target_ids, 'rgbcmykw', iris.target_names):
... plt.scatter(X_pca[y == i, 0], X_pca[y == i, 1],
... c=c, label=label)
<matplotlib.collections.PathCollection ...

18.6. Unsupervised Learning: Dimensionality Reduction and Visualization 624 18.6. Unsupervised Learning: Dimensionality Reduction and Visualization 625
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

The central question is: If our estimator is underperforming, how should we move forward?
• Use simpler or more complicated model?
• Add more features to each observed data point?
• Add more training samples?
The answer is often counter-intuitive. In particular, Sometimes using a more complicated model
will give worse results. Also, Sometimes adding training data will not improve your results.
The ability to determine what steps will improve your model is what separates the successful machine
learning practitioners from the unsuccessful.

Bias-variance trade-off: illustration on a simple regression problem

Code and notebook

Python code and Jupyter notebook for this section are found here

Let us start with a simple 1D regression problem. This will help us to easily visualize the data and the
model, and the results generalize easily to higher-dimensional datasets. We’ll explore a simple linear
regression problem, with sklearn.linear_model.

X = np.c_[0.5, 1].T
fit_transform y = [0.5, 1]
X_test = np.c_[0, 2].T
As TSNE cannot be applied to new data, we need to use its fit_transform method.
Without noise, as linear regression fits the data perfectly

from sklearn import linear_model

regr = linear_model.LinearRegression()
regr.fit(X, y)
sklearn.manifold.TSNE separates quite well the different classes of digits even though it had no access plt.plot(X, y, "o")
to the class information. plt.plot(X_test, regr.predict(X_test))

Exercise: Other dimension reduction of digits

sklearn.manifold has many other non-linear embeddings. Try them out on the digits dataset.
Could you judge their quality without knowing the labels y?
>>> from sklearn.datasets import load_digits
>>> digits = load_digits()
>>> # ...

18.7 Parameter selection, Validation, and Testing

18.7.1 Hyperparameters, Over-fitting, and Under-fitting [<matplotlib.lines.Line2D object at 0x7f246420acf0>]

In real life situation, we have noise (e.g. measurement noise) in our data:
ã See also
rng = np.random.default_rng(27446968)
This section is adapted from Andrew Ng’s excellent Coursera course
for _ in range(6):
noisy_X = X + np.random.normal(loc=0, scale=0.1, size=X.shape)
The issues associated with validation and cross-validation are some of the most important aspects of the plt.plot(noisy_X, y, "o")
practice of machine learning. Selecting the optimal model for your data is vital, and is a piece of the regr.fit(noisy_X, y)
problem that is not often appreciated by machine learning practitioners. plt.plot(X_test, regr.predict(X_test))

18.7. Parameter selection, Validation, and Testing 626 18.7. Parameter selection, Validation, and Testing 627
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

the polynomial) to give the best fit.

On a given data, let us fit a simple polynomial regression model with varying degrees:

As we can see, our linear model captures and amplifies the noise in the data. It displays a lot of variance.
We can use another linear estimator that uses regularization, the Ridge estimator. This estimator
regularizes the coefficients by shrinking them to zero, under the assumption that very high correlations
are often spurious. The alpha parameter controls the amount of shrinkage used.

regr = linear_model.Ridge(alpha=0.1)
np.random.seed(0)
for _ in range(6):
noisy_X = X + np.random.normal(loc=0, scale=0.1, size=X.shape)
 Tip
plt.plot(noisy_X, y, "o")
regr.fit(noisy_X, y) In the above figure, we see fits for three different values of d. For d = 1, the data is under-fit. This
plt.plot(X_test, regr.predict(X_test)) means that the model is too simplistic: no straight line will ever be a good fit to this data. In
this case, we say that the model suffers from high bias. The model itself is biased, and this will be
plt.show() reflected in the fact that the data is poorly fit. At the other extreme, for d = 6 the data is over-fit.
This means that the model has too many free parameters (6 in this case) which can be adjusted to
perfectly fit the training data. If we add a new point to this plot, though, chances are it will be very
far from the curve representing the degree-6 fit. In this case, we say that the model suffers from high
variance. The reason for the term “high variance” is that if any of the input points are varied slightly,
it could result in a very different model.
In the middle, for d = 2, we have found a good mid-point. It fits the data fairly well, and does not
suffer from the bias and variance problems seen in the figures on either side. What we would like is a
way to quantitatively identify bias and variance, and optimize the metaparameters (in this case, the
polynomial degree d) in order to determine the best algorithm.

Polynomial regression with scikit-learn

As we can see, the estimator displays much less variance. However it systematically under-estimates the A polynomial regression is built by pipelining PolynomialFeatures and a LinearRegression:
coefficient. It displays a biased behavior.
>>> from sklearn.pipeline import make_pipeline
This is a typical example of bias/variance tradeof : non-regularized estimator are not biased, but they >>> from sklearn.preprocessing import PolynomialFeatures
can display a lot of variance. Highly-regularized models have little variance, but high bias. This bias is >>> from sklearn.linear_model import LinearRegression
not necessarily a bad thing: what matters is choosing the tradeoff between bias and variance that leads >>> model = make_pipeline(PolynomialFeatures(degree=2), LinearRegression())
to the best prediction performance. For a specific dataset there is a sweet spot corresponding to the
highest complexity that the data can support, depending on the amount of noise and of observations
available. Validation Curves
Let us create a dataset like in the example above:
18.7.2 Visualizing the Bias/Variance Tradeoff
>>> def generating_func(x, rng, err=0.5):
... return rng.normal(10 - 1. / (x + 0.1), err)
 Tip

Given a particular dataset and a model (e.g. a polynomial), we’d like to understand whether bias >>> # randomly sample more data
(underfit) or variance limits prediction, and how to tune the hyperparameter (here d, the degree of >>> rng = np.random.default_rng(27446968)
(continues on next page)

18.7. Parameter selection, Validation, and Testing 628 18.7. Parameter selection, Validation, and Testing 629
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

>>> x = rng.random(size=200)
>>> y = generating_func(x, err=1., rng=rng)

This figure shows why validation is important. On the left side of the plot, we have very low-degree
polynomial, which under-fit the data. This leads to a low explained variance for both the training set
and the validation set. On the far right side of the plot, we have a very high degree polynomial, which
over-fits the data. This can be seen in the fact that the training explained variance is very high, while
Central to quantify bias and variance of a model is to apply it on test data, sampled from the same on the validation set, it is low. Choosing d around 4 or 5 gets us the best tradeoff.
distribution as the train, but that will capture independent noise:

>>> xtrain, xtest, ytrain, ytest = train_test_split(x, y, test_size=0.4)  Tip

The astute reader will realize that something is amiss here: in the above plot, d = 4 gives the best
Validation curve A validation curve consists in varying a model parameter that controls its complexity results. But in the previous plot, we found that d = 6 vastly over-fits the data. What’s going on
(here the degree of the polynomial) and measures both error of the model on training data, and on test here? The difference is the number of training points used. In the previous example, there were
data (eg with cross-validation). The model parameter is then adjusted so that the test error is minimized: only eight training points. In this example, we have 100. As a general rule of thumb, the more
We use sklearn.model_selection.validation_curve() to compute train and test error, and plot it: training points used, the more complicated model can be used. But how can you determine for a
given model whether more training points will be helpful? A useful diagnostic for this are learning
>>> from sklearn.model_selection import validation_curve curves.

>>> degrees = np.arange(1, 21)

Learning Curves
>>> model = make_pipeline(PolynomialFeatures(), LinearRegression()) A learning curve shows the training and validation score as a function of the number of training points.
Note that when we train on a subset of the training data, the training score is computed using this
>>> # Vary the "degrees" on the pipeline step "polynomialfeatures" subset, not the full training set. This curve gives a quantitative view into how beneficial it will be to
>>> train_scores, validation_scores = validation_curve( add training samples.
... model, x[:, np.newaxis], y,
... param_name='polynomialfeatures__degree',
... param_range=degrees) Questions:

>>> # Plot the mean train score and validation score across folds • As the number of training samples are increased, what do you expect to see for the training
>>> plt.plot(degrees, validation_scores.mean(axis=1), label='cross-validation') score? For the validation score?
[<matplotlib.lines.Line2D object at ...>] • Would you expect the training score to be higher or lower than the validation score? Would
>>> plt.plot(degrees, train_scores.mean(axis=1), label='training') you ever expect this to change?
[<matplotlib.lines.Line2D object at ...>]
>>> plt.legend(loc='best')
scikit-learn provides sklearn.model_selection.learning_curve():
<matplotlib.legend.Legend object at ...>
>>> from sklearn.model_selection import learning_curve
>>> train_sizes, train_scores, validation_scores = learning_curve(
... model, x[:, np.newaxis], y, train_sizes=np.logspace(-1, 0, 20))

>>> # Plot the mean train score and validation score across folds
>>> plt.plot(train_sizes, validation_scores.mean(axis=1), label='cross-validation')
[<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(train_sizes, train_scores.mean(axis=1), label='training')
[<matplotlib.lines.Line2D object at ...>]

18.7. Parameter selection, Validation, and Testing 630 18.7. Parameter selection, Validation, and Testing 631
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Here we show the learning curve for d = 15. From the above discussion, we know that d = 15 is a
high-variance estimator which over-fits the data. This is indicated by the fact that the training score
is much higher than the validation score. As we add more samples to this training set, the training score
will continue to decrease, while the cross-validation error will continue to increase, until they meet in
the middle.

Learning curves that have not yet converged with the full training set indicate a high-
variance, over-fit model.
A high-variance model can be improved by:
• Gathering more training samples.
• Using a less-sophisticated model (i.e. in this case, make d smaller)
• Increasing regularization.
Fig. 5: For a degree=1 model In particular, gathering more features for each sample will not help the results.

Note that the validation score generally increases with a growing training set, while the training score 18.7.3 Summary on model selection
generally decreases with a growing training set. As the training size increases, they will converge to a We’ve seen above that an under-performing algorithm can be due to two possible situations: high
single value. bias (under-fitting) and high variance (over-fitting). In order to evaluate our algorithm, we set aside
From the above discussion, we know that d = 1 is a high-bias estimator which under-fits the data. This a portion of our training data for cross-validation. Using the technique of learning curves, we can
is indicated by the fact that both the training and validation scores are low. When confronted with this train on progressively larger subsets of the data, evaluating the training error and cross-validation error
type of learning curve, we can expect that adding more training data will not help: both lines converge to determine whether our algorithm has high variance or high bias. But what do we do with this
to a relatively low score. information?

High Bias
When the learning curves have converged to a low score, we have a high bias model.
If a model shows high bias, the following actions might help:
A high-bias model can be improved by:
• Add more features. In our example of predicting home prices, it may be helpful to make use of
• Using a more sophisticated model (i.e. in this case, increase d) information such as the neighborhood the house is in, the year the house was built, the size of the
lot, etc. Adding these features to the training and test sets can improve a high-bias estimator
• Gather more features for each sample.
• Use a more sophisticated model. Adding complexity to the model can help improve on bias.
• Decrease regularization in a regularized model.
For a polynomial fit, this can be accomplished by increasing the degree d. Each learning technique
Increasing the number of samples, however, does not improve a high-bias model. has its own methods of adding complexity.
Now let’s look at a high-variance (i.e. over-fit) model: • Use fewer samples. Though this will not improve the classification, a high-bias algorithm can
attain nearly the same error with a smaller training sample. For algorithms which are compu-
tationally expensive, reducing the training sample size can lead to very large improvements in
speed.
• Decrease regularization. Regularization is a technique used to impose simplicity in some ma-
chine learning models, by adding a penalty term that depends on the characteristics of the param-
eters. If a model has high bias, decreasing the effect of regularization can lead to better results.

High Variance
If a model shows high variance, the following actions might help:
• Use fewer features. Using a feature selection technique may be useful, and decrease the over-
fitting of the estimator.
• Use a simpler model. Model complexity and over-fitting go hand-in-hand.
• Use more training samples. Adding training samples can reduce the effect of over-fitting, and
lead to improvements in a high variance estimator.
Fig. 6: For a degree=15 model
• Increase Regularization. Regularization is designed to prevent over-fitting. In a high-variance
model, increasing regularization can lead to better results.

18.7. Parameter selection, Validation, and Testing 632 18.7. Parameter selection, Validation, and Testing 633
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

These choices become very important in real-world situations. For example, due to limited telescope (continued from previous page)
time, astronomers must seek a balance between observing a large number of objects, and observing a plt.legend()
large number of features for each object. Determining which is more important for a particular learning plt.show()
task can inform the observing strategy that the astronomer employs.

18.7.4 A last word of caution: separate validation and test set

Using validation schemes to determine hyper-parameters means that we are fitting the hyper-parameters
to the particular validation set. In the same way that parameters can be over-fit to the training set,
hyperparameters can be over-fit to the validation set. Because of this, the validation error tends to
under-predict the classification error of new data.
For this reason, it is recommended to split the data into three sets:
• The training set, used to train the model (usually ~60% of the data)
• The validation set, used to validate the model (usually ~20% of the data)
• The test set, used to evaluate the expected error of the validated model (usually ~20% of the
data)
Many machine learning practitioners do not separate test set and validation set. But if your goal is to
gauge the error of a model on unknown data, using an independent test set is vital.

18.8 Examples for the scikit-learn chapter

18.8.1 Demo PCA in 2D
Load the iris data

from sklearn import datasets

iris = datasets.load_iris()
X = iris.data
y = iris.target Total running time of the script: (0 minutes 0.102 seconds)

Fit a PCA

from sklearn.decomposition import PCA 18.8.2 Measuring Decision Tree performance

Demonstrates overfit when testing on train set.
pca = PCA(n_components=2, whiten=True)
pca.fit(X) Get the data

from sklearn.datasets import fetch_california_housing

Project the data in 2D

X_pca = pca.transform(X) data = fetch_california_housing(as_frame=True)

Visualize the data Train and test a model

target_ids = range(len(iris.target_names)) from sklearn.tree import DecisionTreeRegressor

import matplotlib.pyplot as plt clf = DecisionTreeRegressor().fit(data.data, data.target)

plt.figure(figsize=(6, 5)) predicted = clf.predict(data.data)

for i, c, label in zip(target_ids, "rgbcmykw", iris.target_names, strict=False): expected = data.target
plt.scatter(X_pca[y == i, 0], X_pca[y == i, 1], c=c, label=label)
(continues on next page)
Plot predicted as a function of expected

18.8. Examples for the scikit-learn chapter 634 18.8. Examples for the scikit-learn chapter 635
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

import matplotlib.pyplot as plt import numpy as np

import matplotlib.pyplot as plt
plt.figure(figsize=(4, 3)) from sklearn.linear_model import LinearRegression
plt.scatter(expected, predicted)
plt.plot([0, 5], [0, 5], "--k") # x from 0 to 30
plt.axis("tight") rng = np.random.default_rng()
plt.xlabel("True price ($100k)") x = 30 * rng.random((20, 1))
plt.ylabel("Predicted price ($100k)")
plt.tight_layout() # y = a*x + b with noise
y = 0.5 * x + 1.0 + rng.normal(size=x.shape)

# create a linear regression model

model = LinearRegression()
model.fit(x, y)

# predict y from the data

x_new = np.linspace(0, 30, 100)
y_new = model.predict(x_new[:, np.newaxis])

# plot the results

plt.figure(figsize=(4, 3))
ax = plt.axes()
ax.scatter(x, y)
ax.plot(x_new, y_new)

ax.set_xlabel("x")
ax.set_ylabel("y")

ax.axis("tight")
Pretty much no errors!
This is too good to be true: we are testing the model on the train data, which is not a measure of
generalization. plt.show()

The results are not valid Total running time of the script: (0 minutes 0.048 seconds)
Total running time of the script: (0 minutes 1.459 seconds)

18.8.4 Plot 2D views of the iris dataset

18.8.3 A simple linear regression Plot a simple scatter plot of 2 features of the iris dataset.
Note that more elaborate visualization of this dataset is detailed in the Statistics in Python chapter.

18.8. Examples for the scikit-learn chapter 636 18.8. Examples for the scikit-learn chapter 637
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Load the iris data

from sklearn import datasets

digits = datasets.load_digits()
# Take the first 500 data points: it's hard to see 1500 points
X = digits.data[:500]
y = digits.target[:500]

Fit and transform with a TSNE

from sklearn.manifold import TSNE

tsne = TSNE(n_components=2, random_state=0)

Project the data in 2D

X_2d = tsne.fit_transform(X)

Visualize the data

target_ids = range(len(digits.target_names))

import matplotlib.pyplot as plt

plt.figure(figsize=(6, 5))
# Load the data colors = "r", "g", "b", "c", "m", "y", "k", "w", "orange", "purple"
from sklearn.datasets import load_iris for i, c, label in zip(target_ids, colors, digits.target_names, strict=True):
plt.scatter(X_2d[y == i, 0], X_2d[y == i, 1], c=c, label=label)
iris = load_iris() plt.legend()
plt.show()
from matplotlib import ticker
import matplotlib.pyplot as plt

# The indices of the features that we are plotting

x_index = 0
y_index = 1

# this formatter will label the colorbar with the correct target names
formatter = ticker.FuncFormatter(lambda i, *args: iris.target_names[int(i)])

plt.figure(figsize=(5, 4))
plt.scatter(iris.data[:, x_index], iris.data[:, y_index], c=iris.target)
plt.colorbar(ticks=[0, 1, 2], format=formatter)
plt.xlabel(iris.feature_names[x_index])
plt.ylabel(iris.feature_names[y_index])

plt.tight_layout()
plt.show()

Total running time of the script: (0 minutes 0.094 seconds)

18.8.5 tSNE to visualize digits

Here we use sklearn.manifold.TSNE to visualize the digits datasets. Indeed, the digits are vectors in
a 8*8 = 64 dimensional space. We want to project them in 2D for visualization. tSNE is often a good
solution, as it groups and separates data points based on their local relationship.

18.8. Examples for the scikit-learn chapter 638 18.8. Examples for the scikit-learn chapter 639
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

We compute the cross-validation score as a function of alpha, the strength of the regularization for Lasso
and Ridge

import numpy as np
import matplotlib.pyplot as plt

alphas = np.logspace(-3, -1, 30)

plt.figure(figsize=(5, 3))

for Model in [Lasso, Ridge]:

scores = [cross_val_score(Model(alpha), X, y, cv=3).mean() for alpha in alphas]
plt.plot(alphas, scores, label=Model.__name__)

plt.legend(loc="lower left")
plt.xlabel("alpha")
plt.ylabel("cross validation score")
plt.tight_layout()
plt.show()

Total running time of the script: (0 minutes 1.236 seconds)

18.8.6 Use the RidgeCV and LassoCV to set the regularization parameter
Load the diabetes dataset

from sklearn.datasets import load_diabetes

Total running time of the script: (0 minutes 0.438 seconds)
data = load_diabetes()
X, y = data.data, data.target
print(X.shape)
18.8.7 Plot variance and regularization in linear models
(442, 10) import numpy as np

Compute the cross-validation score with the default hyper-parameters # Smaller figures
import matplotlib.pyplot as plt
from sklearn.model_selection import cross_val_score
from sklearn.linear_model import Ridge, Lasso plt.rcParams["figure.figsize"] = (3, 2)

for Model in [Ridge, Lasso]: We consider the situation where we have only 2 data point
model = Model()
print(f"{ Model.__name__} : { cross_val_score(model, X, y).mean()} ") X = np.c_[0.5, 1].T
y = [0.5, 1]
X_test = np.c_[0, 2].T
Ridge: 0.410174971340889
Lasso: 0.3375593674654274
Without noise, as linear regression fits the data perfectly

18.8. Examples for the scikit-learn chapter 640 18.8. Examples for the scikit-learn chapter 641
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

from sklearn import linear_model
plt.show()
regr = linear_model.LinearRegression()
regr.fit(X, y)
plt.plot(X, y, "o")
plt.plot(X_test, regr.predict(X_test))

Total running time of the script: (0 minutes 0.107 seconds)

18.8.8 Simple picture of the formal problem of machine learning

[<matplotlib.lines.Line2D object at 0x7f246420acf0>]
This example generates simple synthetic data ploints and shows a separating hyperplane on them.
In real life situation, we have noise (e.g. measurement noise) in our data:

rng = np.random.default_rng(27446968)
for _ in range(6):
noisy_X = X + np.random.normal(loc=0, scale=0.1, size=X.shape)
plt.plot(noisy_X, y, "o")
regr.fit(noisy_X, y)
plt.plot(X_test, regr.predict(X_test))

import numpy as np
import matplotlib.pyplot as plt
As we can see, our linear model captures and amplifies the noise in the data. It displays a lot of variance. from sklearn.linear_model import SGDClassifier
from sklearn.datasets import make_blobs
We can use another linear estimator that uses regularization, the Ridge estimator. This estimator
regularizes the coefficients by shrinking them to zero, under the assumption that very high correlations # we create 50 separable synthetic points
are often spurious. The alpha parameter controls the amount of shrinkage used. X, Y = make_blobs(n_samples=50, centers=2, random_state=0, cluster_std=0.60)
regr = linear_model.Ridge(alpha=0.1)
np.random.seed(0) # fit the model
for _ in range(6): clf = SGDClassifier(loss="hinge", alpha=0.01, fit_intercept=True)
noisy_X = X + np.random.normal(loc=0, scale=0.1, size=X.shape) clf.fit(X, Y)
plt.plot(noisy_X, y, "o")
regr.fit(noisy_X, y) # plot the line, the points, and the nearest vectors to the plane
plt.plot(X_test, regr.predict(X_test)) xx = np.linspace(-1, 5, 10)
yy = np.linspace(-1, 5, 10)
(continues on next page)
(continues on next page)

18.8. Examples for the scikit-learn chapter 642 18.8. Examples for the scikit-learn chapter 643
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

from sklearn import model_selection, datasets, metrics
from sklearn.svm import LinearSVC
X1, X2 = np.meshgrid(xx, yy) from sklearn.naive_bayes import GaussianNB
Z = np.empty(X1.shape) from sklearn.neighbors import KNeighborsClassifier
for (i, j), val in np.ndenumerate(X1):
x1 = val digits = datasets.load_digits()
x2 = X2[i, j] X = digits.data
p = clf.decision_function([[x1, x2]]) y = digits.target
Z[i, j] = p[0] X_train, X_test, y_train, y_test = model_selection.train_test_split(
X, y, test_size=0.25, random_state=0
plt.figure(figsize=(4, 3)) )
ax = plt.axes()
ax.contour( for Model in [LinearSVC, GaussianNB, KNeighborsClassifier]:
X1, X2, Z, [-1.0, 0.0, 1.0], colors="k", linestyles=["dashed", "solid", "dashed"] clf = Model().fit(X_train, y_train)
) y_pred = clf.predict(X_test)
ax.scatter(X[:, 0], X[:, 1], c=Y, cmap="Paired") print(f"{ Model.__name__} : { metrics.f1_score(y_test, y_pred, average='macro')} ")

ax.axis("tight") print("------------------")

# test SVC loss

plt.show() for loss in ["hinge", "squared_hinge"]:
clf = LinearSVC(loss=loss).fit(X_train, y_train)
Total running time of the script: (0 minutes 0.056 seconds) y_pred = clf.predict(X_test)
print(
f"LinearSVC(loss='{ loss} '): { metrics.f1_score(y_test, y_pred, average='macro
˓→')} "
18.8.9 Compare classifiers on the digits data )
Compare the performance of a variety of classifiers on a test set for the digits data.
print("-------------------")
LinearSVC: 0.9344942114287969
GaussianNB: 0.8332741681010102 # test the number of neighbors
KNeighborsClassifier: 0.9804562804949924 for n_neighbors in range(1, 11):
------------------ clf = KNeighborsClassifier(n_neighbors=n_neighbors).fit(X_train, y_train)
/opt/hostedtoolcache/Python/3.12.9/x64/lib/python3.12/site-packages/sklearn/svm/_base. y_pred = clf.predict(X_test)
˓→py:1249: ConvergenceWarning: Liblinear failed to converge, increase the number of␣ print(
˓→iterations. f"KNeighbors(n_neighbors={ n_neighbors} ): { metrics.f1_score(y_test, y_pred,␣
warnings.warn( ˓→average='macro')} "

LinearSVC(loss='hinge'): 0.9294570108037394 )
LinearSVC(loss='squared_hinge'): 0.9344942114287969
------------------- Total running time of the script: (0 minutes 0.261 seconds)
KNeighbors(n_neighbors=1): 0.9913675218842191
KNeighbors(n_neighbors=2): 0.9848442068835102
KNeighbors(n_neighbors=3): 0.9867753449543099
KNeighbors(n_neighbors=4): 0.9803719053818863 18.8.10 Plot fitting a 9th order polynomial
KNeighbors(n_neighbors=5): 0.9804562804949924
Fits data generated from a 9th order polynomial with model of 4th order and 9th order polynomials, to
KNeighbors(n_neighbors=6): 0.9757924194139573
demonstrate that often simpler models are to be preferred
KNeighbors(n_neighbors=7): 0.9780645792142071
KNeighbors(n_neighbors=8): 0.9780645792142071 import numpy as np
KNeighbors(n_neighbors=9): 0.9780645792142071 import matplotlib.pyplot as plt
KNeighbors(n_neighbors=10): 0.9755550897728812 from matplotlib.colors import ListedColormap

from sklearn import linear_model

# Create color maps for 3-class classification problem, as with iris

cmap_light = ListedColormap(["#FFAAAA", "#AAFFAA", "#AAAAFF"])
cmap_bold = ListedColormap(["#FF0000", "#00FF00", "#0000FF"])

(continues on next page)

18.8. Examples for the scikit-learn chapter 644 18.8. Examples for the scikit-learn chapter 645
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

regr.fit(X, y)
rng = np.random.default_rng(27446968) plt.plot(x_test, regr.predict(X_test), label="9th order")
x = 2 * rng.random(100) - 1
plt.legend(loc="best")
f = lambda t: 1.2 * t**2 + 0.1 * t**3 - 0.4 * t**5 - 0.5 * t**9 plt.axis("tight")
y = f(x) + 0.4 * rng.normal(size=100) plt.title("Fitting a 4th and a 9th order polynomial")

x_test = np.linspace(-1, 1, 100)

The data

plt.figure(figsize=(6, 4))
plt.scatter(x, y, s=4)

Text(0.5, 1.0, 'Fitting a 4th and a 9th order polynomial')

Ground truth

plt.figure(figsize=(6, 4))
plt.scatter(x, y, s=4)
plt.plot(x_test, f(x_test), label="truth")
<matplotlib.collections.PathCollection object at 0x7f2463f8a120> plt.axis("tight")
plt.title("Ground truth (9th order polynomial)")
Fitting 4th and 9th order polynomials
plt.show()
For this we need to engineer features: the n_th powers of x:

plt.figure(figsize=(6, 4))
plt.scatter(x, y, s=4)

X = np.array([x**i for i in range(5)]).T

X_test = np.array([x_test**i for i in range(5)]).T
regr = linear_model.LinearRegression()
regr.fit(X, y)
plt.plot(x_test, regr.predict(X_test), label="4th order")

X = np.array([x**i for i in range(10)]).T

X_test = np.array([x_test**i for i in range(10)]).T
regr = linear_model.LinearRegression()
(continues on next page)

18.8. Examples for the scikit-learn chapter 646 18.8. Examples for the scikit-learn chapter 647
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Print the join histogram for each feature

for index, feature_name in enumerate(data.feature_names):

plt.figure(figsize=(4, 3))
plt.scatter(data.data[feature_name], data.target)
plt.ylabel("Price", size=15)
Total running time of the script: (0 minutes 0.174 seconds) plt.xlabel(feature_name, size=15)
plt.tight_layout()

18.8.11 A simple regression analysis on the California housing data

Here we perform a simple regression analysis on the California housing data, exploring two types of
regressors.

from sklearn.datasets import fetch_california_housing

data = fetch_california_housing(as_frame=True)

Print a histogram of the quantity to predict: price

import matplotlib.pyplot as plt

plt.figure(figsize=(4, 3))
plt.hist(data.target)
plt.xlabel("price ($100k)")
plt.ylabel("count")
plt.tight_layout() •

18.8. Examples for the scikit-learn chapter 648 18.8. Examples for the scikit-learn chapter 649
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

• •

• •

• •

18.8. Examples for the scikit-learn chapter 650 18.8. Examples for the scikit-learn chapter 651
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Prediction with gradient boosted tree

from sklearn.ensemble import GradientBoostingRegressor

clf = GradientBoostingRegressor()
clf.fit(X_train, y_train)

predicted = clf.predict(X_test)
expected = y_test

plt.figure(figsize=(4, 3))
plt.scatter(expected, predicted)
plt.plot([0, 5], [0, 5], "--k")
plt.axis("tight")
plt.xlabel("True price ($100k)")
plt.ylabel("Predicted price ($100k)")
plt.tight_layout()
•
Simple prediction

from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(data.data, data.target)

from sklearn.linear_model import LinearRegression

clf = LinearRegression()
clf.fit(X_train, y_train)
predicted = clf.predict(X_test)
expected = y_test

plt.figure(figsize=(4, 3))
plt.scatter(expected, predicted)
plt.plot([0, 8], [0, 8], "--k")
plt.axis("tight")
plt.xlabel("True price ($100k)")
plt.ylabel("Predicted price ($100k)") Print the error rate
plt.tight_layout()
import numpy as np

print(f"RMS: { np.sqrt(np.mean((predicted - expected) ** 2))!r} ")

plt.show()

RMS: np.float64(0.5314909993118918)

Total running time of the script: (0 minutes 4.743 seconds)

18.8.12 Nearest-neighbor prediction on iris

Plot the decision boundary of nearest neighbor decision on iris, first with a single nearest neighbor, and
then using 3 nearest neighbors.
import numpy as np
import matplotlib.pyplot as plt
from sklearn import neighbors, datasets
(continues on next page)

18.8. Examples for the scikit-learn chapter 652 18.8. Examples for the scikit-learn chapter 653
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

from matplotlib.colors import ListedColormap

# Create color maps for 3-class classification problem, as with iris

cmap_light = ListedColormap(["#FFAAAA", "#AAFFAA", "#AAAAFF"])
cmap_bold = ListedColormap(["#FF0000", "#00FF00", "#0000FF"])

iris = datasets.load_iris()
X = iris.data[:, :2] # we only take the first two features. We could
# avoid this ugly slicing by using a two-dim dataset
y = iris.target

knn = neighbors.KNeighborsClassifier(n_neighbors=1)
knn.fit(X, y)

x_min, x_max = X[:, 0].min() - 0.1, X[:, 0].max() + 0.1

y_min, y_max = X[:, 1].min() - 0.1, X[:, 1].max() + 0.1
xx, yy = np.meshgrid(np.linspace(x_min, x_max, 100), np.linspace(y_min, y_max, 100))
Z = knn.predict(np.c_[xx.ravel(), yy.ravel()])

Put the result into a color plot

Z = Z.reshape(xx.shape)
plt.figure()
plt.pcolormesh(xx, yy, Z, cmap=cmap_light)

# Plot also the training points

plt.scatter(X[:, 0], X[:, 1], c=y, cmap=cmap_bold)
plt.xlabel("sepal length (cm)")
plt.ylabel("sepal width (cm)") (np.float64(4.180808080808081), np.float64(8.019191919191918), np.float64(1.
plt.axis("tight") ˓→8868686868686868), np.float64(4.513131313131313))

And now, redo the analysis with 3 neighbors

knn = neighbors.KNeighborsClassifier(n_neighbors=3)
knn.fit(X, y)

Z = knn.predict(np.c_[xx.ravel(), yy.ravel()])

# Put the result into a color plot

Z = Z.reshape(xx.shape)
plt.figure()
plt.pcolormesh(xx, yy, Z, cmap=cmap_light)

# Plot also the training points

plt.scatter(X[:, 0], X[:, 1], c=y, cmap=cmap_bold)
plt.xlabel("sepal length (cm)")
plt.ylabel("sepal width (cm)")
plt.axis("tight")

plt.show()

18.8. Examples for the scikit-learn chapter 654 18.8. Examples for the scikit-learn chapter 655
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

Total running time of the script: (0 minutes 0.776 seconds)

18.8.13 Simple visualization and classification of the digits dataset

Plot the first few samples of the digits dataset and a 2D representation built using PCA, then do a simple
classification

from sklearn.datasets import load_digits Plot a projection on the 2 first principal axis
digits = load_digits() plt.figure()

from sklearn.decomposition import PCA

Plot the data: images of digits
Each data in a 8x8 image pca = PCA(n_components=2)
proj = pca.fit_transform(digits.data)
import matplotlib.pyplot as plt plt.scatter(proj[:, 0], proj[:, 1], c=digits.target, cmap="Paired")
plt.colorbar()
fig = plt.figure(figsize=(6, 6)) # figure size in inches
fig.subplots_adjust(left=0, right=1, bottom=0, top=1, hspace=0.05, wspace=0.05)

for i in range(64):
ax = fig.add_subplot(8, 8, i + 1, xticks=[], yticks=[])
ax.imshow(digits.images[i], cmap="binary", interpolation="nearest")
# label the image with the target value
ax.text(0, 7, str(digits.target[i]))

18.8. Examples for the scikit-learn chapter 656 18.8. Examples for the scikit-learn chapter 657
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

if predicted[i] == expected[i]:
ax.text(0, 7, str(predicted[i]), color="green")
else:
ax.text(0, 7, str(predicted[i]), color="red")

<matplotlib.colorbar.Colorbar object at 0x7f2464e7d3d0>

Classify with Gaussian naive Bayes

from sklearn.naive_bayes import GaussianNB

from sklearn.model_selection import train_test_split

# split the data into training and validation sets

X_train, X_test, y_train, y_test = train_test_split(digits.data, digits.target)

# train the model

clf = GaussianNB()
clf.fit(X_train, y_train)

# use the model to predict the labels of the test data Quantify the performance
predicted = clf.predict(X_test)
expected = y_test First print the number of correct matches

matches = predicted == expected

# Plot the prediction print(matches.sum())
fig = plt.figure(figsize=(6, 6)) # figure size in inches
fig.subplots_adjust(left=0, right=1, bottom=0, top=1, hspace=0.05, wspace=0.05)
395
# plot the digits: each image is 8x8 pixels
for i in range(64): The total number of data points
ax = fig.add_subplot(8, 8, i + 1, xticks=[], yticks=[])
print(len(matches))
ax.imshow(X_test.reshape(-1, 8, 8)[i], cmap="binary", interpolation="nearest")

# label the image with the target value 450

(continues on next page)

18.8. Examples for the scikit-learn chapter 658 18.8. Examples for the scikit-learn chapter 659
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

And now, the ration of correct predictions fetch_lfw_people(). However, this is a relatively large download (~200MB) so we will do the tutorial
on a simpler, less rich dataset. Feel free to explore the LFW dataset.
matches.sum() / float(len(matches))
from sklearn import datasets
np.float64(0.8777777777777778)
faces = datasets.fetch_olivetti_faces()
Print the classification report faces.data.shape

from sklearn import metrics downloading Olivetti faces from https://ndownloader.figshare.com/files/5976027 to /

˓→home/runner/scikit_learn_data
print(metrics.classification_report(expected, predicted))
(400, 4096)
precision recall f1-score support
Let’s visualize these faces to see what we’re working with
0 0.97 0.95 0.96 37
1 0.83 0.85 0.84 41 import matplotlib.pyplot as plt
2 0.89 0.84 0.86 49
3 0.93 0.83 0.88 47 fig = plt.figure(figsize=(8, 6))
4 0.93 0.90 0.92 42 # plot several images
5 0.89 0.95 0.92 42 for i in range(15):
6 0.98 0.97 0.97 60 ax = fig.add_subplot(3, 5, i + 1, xticks=[], yticks=[])
7 0.81 0.98 0.88 47 ax.imshow(faces.images[i], cmap="bone")
8 0.65 0.87 0.75 39
9 0.97 0.63 0.76 46

accuracy 0.88 450

macro avg 0.89 0.88 0.87 450
weighted avg 0.89 0.88 0.88 450

Print the confusion matrix

print(metrics.confusion_matrix(expected, predicted))

plt.show()

[[35 0 0 0 1 0 0 1 0 0]
[ 0 35 0 0 0 0 1 1 4 0]
[ 0 1 41 0 0 0 0 0 7 0]
[ 0 0 2 39 0 1 0 2 2 1]
[ 0 1 0 0 38 0 0 2 1 0]
[ 0 0 0 0 1 40 0 1 0 0]
[ 0 0 1 0 1 0 58 0 0 0]
[ 0 0 0 0 0 1 0 46 0 0]
[ 0 2 0 1 0 1 0 1 34 0]
[ 1 3 2 2 0 2 0 3 4 29]]

Total running time of the script: (0 minutes 1.450 seconds)

18.8.14 The eigenfaces example: chaining PCA and SVMs

The goal of this example is to show how an unsupervised method and a supervised one can be chained
for better prediction. It starts with a didactic but lengthy way of doing things, and finishes with the  Tip
idiomatic approach to pipelining in scikit-learn.
Note is that these faces have already been localized and scaled to a common size. This is an important
Here we’ll take a look at a simple facial recognition example. Ideally, we would use a dataset con-
preprocessing piece for facial recognition, and is a process that can require a large collection of training
sisting of a subset of the Labeled Faces in the Wild data that is available with sklearn.datasets.
data. This can be done in scikit-learn, but the challenge is gathering a sufficient amount of training

18.8. Examples for the scikit-learn chapter 660 18.8. Examples for the scikit-learn chapter 661
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

data for the algorithm to work. Fortunately, this piece is common enough that it has been done. One
good resource is OpenCV, the Open Computer Vision Library.

We’ll perform a Support Vector classification of the images. We’ll do a typical train-test split on the
images:

from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(

faces.data, faces.target, random_state=0
)

print(X_train.shape, X_test.shape)

(300, 4096) (100, 4096)

Preprocessing: Principal Component Analysis

1850 dimensions is a lot for SVM. We can use PCA to reduce these 1850 features to a manageable size,
while maintaining most of the information in the dataset.

from sklearn import decomposition

pca = decomposition.PCA(n_components=150, whiten=True)

pca.fit(X_train)

One interesting part of PCA is that it computes the “mean” face, which can be interesting to examine:

plt.imshow(pca.mean_.reshape(faces.images[0].shape), cmap="bone") <matplotlib.image.AxesImage object at 0x7f2462efe480>

The principal components measure deviations about this mean along orthogonal axes.

print(pca.components_.shape)

(150, 4096)

It is also interesting to visualize these principal components:

fig = plt.figure(figsize=(16, 6))

for i in range(30):
ax = fig.add_subplot(3, 10, i + 1, xticks=[], yticks=[])
ax.imshow(pca.components_[i].reshape(faces.images[0].shape), cmap="bone")

18.8. Examples for the scikit-learn chapter 662 18.8. Examples for the scikit-learn chapter 663
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

The components (“eigenfaces”) are ordered by their importance from top-left to bottom-right. We
see that the first few components seem to primarily take care of lighting conditions; the remaining
components pull out certain identifying features: the nose, eyes, eyebrows, etc.
With this projection computed, we can now project our original training and test data onto the PCA
basis:

X_train_pca = pca.transform(X_train)
X_test_pca = pca.transform(X_test)
print(X_train_pca.shape)

(300, 150)

print(X_test_pca.shape)

(100, 150)

These projected components correspond to factors in a linear combination of component images such
that the combination approaches the original face.

Doing the Learning: Support Vector Machines

Now we’ll perform support-vector-machine classification on this reduced dataset:

from sklearn import svm

clf = svm.SVC(C=5.0, gamma=0.001)

clf.fit(X_train_pca, y_train)

Finally, we can evaluate how well this classification did. First, we might plot a few of the test-cases with The classifier is correct on an impressive number of images given the simplicity of its learning model!
the labels learned from the training set: Using a linear classifier on 150 features derived from the pixel-level data, the algorithm correctly identifies
a large number of the people in the images.
import numpy as np
Again, we can quantify this effectiveness using one of several measures from sklearn.metrics. First we
fig = plt.figure(figsize=(8, 6)) can do the classification report, which shows the precision, recall and other measures of the “goodness”
for i in range(15): of the classification:
ax = fig.add_subplot(3, 5, i + 1, xticks=[], yticks=[])
ax.imshow(X_test[i].reshape(faces.images[0].shape), cmap="bone") from sklearn import metrics
y_pred = clf.predict(X_test_pca[i, np.newaxis])[0]
color = "black" if y_pred == y_test[i] else "red" y_pred = clf.predict(X_test_pca)
ax.set_title(y_pred, fontsize="small", color=color) print(metrics.classification_report(y_test, y_pred))

precision recall f1-score support

0 1.00 0.50 0.67 6

1 1.00 1.00 1.00 4
2 0.50 1.00 0.67 2
3 1.00 1.00 1.00 1
4 0.33 1.00 0.50 1
5 1.00 1.00 1.00 5
6 1.00 1.00 1.00 4
7 1.00 0.67 0.80 3
9 1.00 1.00 1.00 1
10 1.00 1.00 1.00 4
11 1.00 1.00 1.00 1
12 0.67 1.00 0.80 2
13 1.00 1.00 1.00 3
14 1.00 1.00 1.00 5
(continues on next page)

18.8. Examples for the scikit-learn chapter 664 18.8. Examples for the scikit-learn chapter 665
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

15 1.00 1.00 1.00 3 )
17 1.00 1.00 1.00 6
19 1.00 1.00 1.00 4 clf.fit(X_train, y_train)
20 1.00 1.00 1.00 1
21 1.00 1.00 1.00 1 y_pred = clf.predict(X_test)
22 1.00 1.00 1.00 2 print(metrics.confusion_matrix(y_pred, y_test))
23 1.00 1.00 1.00 1 plt.show()
24 1.00 1.00 1.00 2
25 1.00 0.50 0.67 2 [[4 0 0 ... 0 0 0]
26 1.00 0.75 0.86 4 [0 4 0 ... 0 0 0]
27 1.00 1.00 1.00 1 [0 0 1 ... 0 0 0]
28 0.67 1.00 0.80 2 ...
29 1.00 1.00 1.00 3 [1 0 0 ... 3 0 0]
30 1.00 1.00 1.00 4 [0 0 0 ... 0 1 0]
31 1.00 1.00 1.00 3 [0 0 0 ... 0 0 3]]
32 1.00 1.00 1.00 3
33 1.00 1.00 1.00 2
34 1.00 1.00 1.00 3 A Note on Facial Recognition
35 1.00 1.00 1.00 1
36 1.00 1.00 1.00 3 Here we have used PCA “eigenfaces” as a pre-processing step for facial recognition. The reason we
37 1.00 1.00 1.00 3 chose this is because PCA is a broadly-applicable technique, which can be useful for a wide array of
38 1.00 1.00 1.00 1 data types. Research in the field of facial recognition in particular, however, has shown that other more
39 1.00 1.00 1.00 3 specific feature extraction methods are can be much more effective.
Total running time of the script: (0 minutes 4.396 seconds)
accuracy 0.94 100
macro avg 0.95 0.96 0.94 100
weighted avg 0.97 0.94 0.94 100
18.8.15 Example of linear and non-linear models
Another interesting metric is the confusion matrix, which indicates how often any two items are mixed- This is an example plot from the tutorial which accompanies an explanation of the support vector
up. The confusion matrix of a perfect classifier would only have nonzero entries on the diagonal, with machine GUI.
zeros on the off-diagonal:
import numpy as np
print(metrics.confusion_matrix(y_test, y_pred)) import matplotlib.pyplot as plt

[[3 0 0 ... 0 0 0] from sklearn import svm

[0 4 0 ... 0 0 0]
[0 0 2 ... 0 0 0]
... rng = np.random.default_rng(27446968)
[0 0 0 ... 3 0 0]
[0 0 0 ... 0 1 0] data that is linearly separable
[0 0 0 ... 0 0 3]]
def linear_model(rseed=42, n_samples=30):
"Generate data according to a linear model"
Pipelining np.random.seed(rseed)
Above we used PCA as a pre-processing step before applying our support vector machine classifier.
Plugging the output of one estimator directly into the input of a second estimator is a commonly used data = np.random.normal(0, 10, (n_samples, 2))
pattern; for this reason scikit-learn provides a Pipeline object which automates this process. The above data[: n_samples // 2] -= 15
problem can be re-expressed as a pipeline as follows: data[n_samples // 2 :] += 15

from sklearn.pipeline import Pipeline labels = np.ones(n_samples)

labels[: n_samples // 2] = -1
clf = Pipeline(
[ return data, labels
("pca", decomposition.PCA(n_components=150, whiten=True)),
("svm", svm.LinearSVC(C=1.0)),
] X, y = linear_model()
(continues on next page) (continues on next page)

18.8. Examples for the scikit-learn chapter 666 18.8. Examples for the scikit-learn chapter 667
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

def nonlinear_model(rseed=27446968, n_samples=30):
clf = svm.SVC(kernel="linear") rng = np.random.default_rng(rseed)
clf.fit(X, y)
radius = 40 * rng.random(n_samples)
plt.figure(figsize=(6, 4)) far_pts = radius > 20
ax = plt.subplot(111, xticks=[], yticks=[]) radius[far_pts] *= 1.2
ax.scatter(X[:, 0], X[:, 1], c=y, cmap="bone") radius[~far_pts] *= 1.1

ax.scatter( theta = rng.random(n_samples) * np.pi * 2

clf.support_vectors_[:, 0],
clf.support_vectors_[:, 1], data = np.empty((n_samples, 2))
s=80, data[:, 0] = radius * np.cos(theta)
edgecolors="k", data[:, 1] = radius * np.sin(theta)
facecolors="none",
) labels = np.ones(n_samples)
labels[far_pts] = -1
delta = 1
y_min, y_max = -50, 50 return data, labels
x_min, x_max = -50, 50
x = np.arange(x_min, x_max + delta, delta)
y = np.arange(y_min, y_max + delta, delta) X, y = nonlinear_model()
X1, X2 = np.meshgrid(x, y) clf = svm.SVC(kernel="rbf", gamma=0.001, coef0=0, degree=3)
Z = clf.decision_function(np.c_[X1.ravel(), X2.ravel()]) clf.fit(X, y)
Z = Z.reshape(X1.shape)
plt.figure(figsize=(6, 4))
ax.contour( ax = plt.subplot(1, 1, 1, xticks=[], yticks=[])
X1, X2, Z, [-1.0, 0.0, 1.0], colors="k", linestyles=["dashed", "solid", "dashed"] ax.scatter(X[:, 0], X[:, 1], c=y, cmap="bone", zorder=2)
)
ax.scatter(
clf.support_vectors_[:, 0],
clf.support_vectors_[:, 1],
s=80,
edgecolors="k",
facecolors="none",
)

delta = 1
y_min, y_max = -50, 50
x_min, x_max = -50, 50
x = np.arange(x_min, x_max + delta, delta)
y = np.arange(y_min, y_max + delta, delta)
X1, X2 = np.meshgrid(x, y)
Z = clf.decision_function(np.c_[X1.ravel(), X2.ravel()])
Z = Z.reshape(X1.shape)

ax.contour(
X1,
X2,
Z,
[-1.0, 0.0, 1.0],
colors="k",
linestyles=["dashed", "solid", "dashed"],
zorder=1,
)
<matplotlib.contour.QuadContourSet object at 0x7f2468789c70>
plt.show()
data with a non-linear separation

18.8. Examples for the scikit-learn chapter 668 18.8. Examples for the scikit-learn chapter 669
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

titles = ["d = 1 (under-fit; high bias)", "d = 2", "d = 6 (over-fit; high variance)"]
degrees = [1, 2, 6]

fig = plt.figure(figsize=(9, 3.5))

fig.subplots_adjust(left=0.06, right=0.98, bottom=0.15, top=0.85, wspace=0.05)

for i, d in enumerate(degrees):
ax = fig.add_subplot(131 + i, xticks=[], yticks=[])
ax.scatter(x, y, marker="x", c="k", s=50)

model = make_pipeline(PolynomialFeatures(d), LinearRegression())

model.fit(x[:, np.newaxis], y)
ax.plot(x_test, model.predict(x_test[:, np.newaxis]), "-b")

ax.set_xlim(-0.2, 1.2)
ax.set_ylim(0, 12)
ax.set_xlabel("house size")
if i == 0:
ax.set_ylabel("price")

ax.set_title(titles[i])

Total running time of the script: (0 minutes 0.069 seconds)

18.8.16 Bias and variance of polynomial fit

Demo overfitting, underfitting, and validation and learning curves with polynomial regression.
Fit polynomes of different degrees to a dataset: for too small a degree, the model underfits, while for too
large a degree, it overfits.

import numpy as np
import matplotlib.pyplot as plt

def generating_func(x, rng=None, error=0.5):

rng = np.random.default_rng(rng) Generate a larger dataset
return rng.normal(10 - 1.0 / (x + 0.1), error)
from sklearn.model_selection import train_test_split
A polynomial regression
n_samples = 200
from sklearn.pipeline import make_pipeline test_size = 0.4
from sklearn.linear_model import LinearRegression error = 1.0
from sklearn.preprocessing import PolynomialFeatures
# randomly sample the data
A simple figure to illustrate the problem x = rng.random(n_samples)
y = generating_func(x, rng=rng, error=error)
n_samples = 8
# split into training, validation, and testing sets.
rng = np.random.default_rng(27446968) x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=test_size)
x = 10 ** np.linspace(-2, 0, n_samples)
y = generating_func(x, rng=rng) # show the training and validation sets
plt.figure(figsize=(6, 4))
x_test = np.linspace(-0.2, 1.2, 1000) plt.scatter(x_train, y_train, color="red", label="Training set")
(continues on next page) (continues on next page)

18.8. Examples for the scikit-learn chapter 670 18.8. Examples for the scikit-learn chapter 671
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

plt.scatter(x_test, y_test, color="blue", label="Test set") plt.ylabel("explained variance")
plt.title("The data") plt.title("Validation curve")
plt.legend(loc="best") plt.tight_layout()

<matplotlib.legend.Legend object at 0x7f24687889b0>

Learning curves
Plot a validation curve Plot train and test error with an increasing number of samples

from sklearn.model_selection import validation_curve # A learning curve for d=1, 5, 15

for d in [1, 5, 15]:
degrees = list(range(1, 21)) model = make_pipeline(PolynomialFeatures(degree=d), LinearRegression())

model = make_pipeline(PolynomialFeatures(), LinearRegression()) from sklearn.model_selection import learning_curve

# The parameter to vary is the "degrees" on the pipeline step train_sizes, train_scores, validation_scores = learning_curve(
# "polynomialfeatures" model, x[:, np.newaxis], y, train_sizes=np.logspace(-1, 0, 20)
train_scores, validation_scores = validation_curve( )
model,
x[:, np.newaxis], # Plot the mean train error and validation error across folds
y, plt.figure(figsize=(6, 4))
param_name="polynomialfeatures__degree", plt.plot(
param_range=degrees, train_sizes, validation_scores.mean(axis=1), lw=2, label="cross-validation"
) )
plt.plot(train_sizes, train_scores.mean(axis=1), lw=2, label="training")
# Plot the mean train error and validation error across folds plt.ylim(ymin=-0.1, ymax=1)
plt.figure(figsize=(6, 4))
plt.plot(degrees, validation_scores.mean(axis=1), lw=2, label="cross-validation") plt.legend(loc="best")
plt.plot(degrees, train_scores.mean(axis=1), lw=2, label="training") plt.xlabel("number of train samples")
plt.ylabel("explained variance")
plt.legend(loc="best") plt.title(f"Learning curve (degree={ d} )")
plt.xlabel("degree of fit") plt.tight_layout()
(continues on next page) (continues on next page)

18.8. Examples for the scikit-learn chapter 672 18.8. Examples for the scikit-learn chapter 673
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

plt.show()

•
Total running time of the script: (0 minutes 1.779 seconds)

• 18.8.17 Tutorial Diagrams

This script plots the flow-charts used in the scikit-learn tutorials.

18.8. Examples for the scikit-learn chapter 674 18.8. Examples for the scikit-learn chapter 675
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page)

ax = plt.axes((0, 0, 1, 1), xticks=[], yticks=[], frameon=False)
ax.set_xlim(0, 9)
ax.set_ylim(0, 6)

patches = [
Rectangle((0.3, 3.6), 1.5, 1.8, zorder=1, fc=box_bg),
Rectangle((0.5, 3.8), 1.5, 1.8, zorder=2, fc=box_bg),
Rectangle((0.7, 4.0), 1.5, 1.8, zorder=3, fc=box_bg),
Rectangle((2.9, 3.6), 0.2, 1.8, fc=box_bg),
Rectangle((3.1, 3.8), 0.2, 1.8, fc=box_bg),
Rectangle((3.3, 4.0), 0.2, 1.8, fc=box_bg),
Rectangle((0.3, 0.2), 1.5, 1.8, fc=box_bg),
Rectangle((2.9, 0.2), 0.2, 1.8, fc=box_bg),
Circle((5.5, 3.5), 1.0, fc=box_bg),
Polygon([[5.5, 1.7], [6.1, 1.1], [5.5, 0.5], [4.9, 1.1]], fc=box_bg),
FancyArrow(
2.3, 4.6, 0.35, 0, fc=arrow1, width=0.25, head_width=0.5, head_length=0.2
),
FancyArrow(
3.75, 4.2, 0.5, -0.2, fc=arrow1, width=0.25, head_width=0.5, head_
˓→length=0.2

),
• FancyArrow(
5.5, 2.4, 0, -0.4, fc=arrow1, width=0.25, head_width=0.5, head_length=0.2
),
FancyArrow(
2.0, 1.1, 0.5, 0, fc=arrow2, width=0.25, head_width=0.5, head_length=0.2
),
FancyArrow(
3.3, 1.1, 1.3, 0, fc=arrow2, width=0.25, head_width=0.5, head_length=0.2
),
FancyArrow(
6.2, 1.1, 0.8, 0, fc=arrow2, width=0.25, head_width=0.5, head_length=0.2
),
]

if supervised:
patches += [
Rectangle((0.3, 2.4), 1.5, 0.5, zorder=1, fc=box_bg),
Rectangle((0.5, 2.6), 1.5, 0.5, zorder=2, fc=box_bg),
Rectangle((0.7, 2.8), 1.5, 0.5, zorder=3, fc=box_bg),
FancyArrow(
2.3, 2.9, 2.0, 0, fc=arrow1, width=0.25, head_width=0.5, head_
˓→length=0.2

),
Rectangle((7.3, 0.85), 1.5, 0.5, fc=box_bg),
• ]
else:
import numpy as np patches += [Rectangle((7.3, 0.2), 1.5, 1.8, fc=box_bg)]
import matplotlib.pyplot as plt
from matplotlib.patches import Circle, Rectangle, Polygon, Arrow, FancyArrow for p in patches:
ax.add_patch(p)

def create_base(box_bg="#CCCCCC", arrow1="#88CCFF", arrow2="#88FF88",␣ plt.text(

˓→supervised=True): 1.45,
fig = plt.figure(figsize=(9, 6), facecolor="w") 4.9,
(continues on next page) (continues on next page)

18.8. Examples for the scikit-learn chapter 676 18.8. Examples for the scikit-learn chapter 677
 Scientific Python Lectures, Edition 2025.1rc0.dev0 Scientific Python Lectures, Edition 2025.1rc0.dev0

(continued from previous page) (continued from previous page)

"Training\nText,\nDocuments,\nImages,\netc.", fontdict=fontdict,
ha="center", rotation=20,
va="center", ha="left",
fontsize=14, va="bottom",
) )
plt.text(
plt.text(3.6, 4.9, "Feature\nVectors", ha="left", va="center", fontsize=14) 3.7,
3.2,
plt.text( "clf.fit(X, y)",
5.5, 3.5, "Machine\nLearning\nAlgorithm", ha="center", va="center",␣ fontdict=fontdict,
˓→fontsize=14 rotation=20,
) ha="left",
va="bottom",
plt.text( )
1.05, plt.text(
1.1, 1.7,
"New Text,\nDocument,\nImage,\netc.", 1.5,
ha="center", "X_new = vec.transform(input)",
va="center", fontdict=fontdict,
fontsize=14, rotation=20,
) ha="left",
va="bottom",
plt.text(3.3, 1.7, "Feature\nVector", ha="left", va="center", fontsize=14) )
plt.text(
plt.text(5.5, 1.1, "Predictive\nModel", ha="center", va="center", fontsize=12) 6.1,
1.5,
if supervised: "y_new = clf.predict(X_new)",
plt.text(1.45, 3.05, "Labels", ha="center", va="center", fontsize=14) fontdict=fontdict,
rotation=20,
plt.text(8.05, 1.1, "Expected\nLabel", ha="center", va="center", fontsize=14) ha="left",
plt.text( va="bottom",
8.8, 5.8, "Supervised Learning Model", ha="right", va="top", fontsize=18 )
)

else: def plot_unsupervised_chart():

plt.text( create_base(supervised=False)
8.05,
1.1,
"Likelihood\nor Cluster ID\nor Better\nRepresentation", if __name__ == "__main__":
ha="center", plot_supervised_chart(False)
va="center", plot_supervised_chart(True)
fontsize=12, plot_unsupervised_chart()
) plt.show()
plt.text(
8.8, 5.8, "Unsupervised Learning Model", ha="right", va="top", fontsize=18 Total running time of the script: (0 minutes 0.194 seconds)
)

def plot_supervised_chart(annotate=False): ã See also

create_base(supervised=True)
if annotate: Going further
fontdict = {"color": "r", "weight": "bold", "size": 14}
• The documentation of scikit-learn is very complete and didactic.
plt.text(
1.9, • Introduction to Machine Learning with Python, by Sarah Guido, Andreas Müller (notebooks
4.55, available here).
"X = vec.fit_transform(input)",
(continues on next page)

18.8. Examples for the scikit-learn chapter 678 18.8. Examples for the scikit-learn chapter 679
 Scientific Python Lectures, Edition 2025.1rc0.dev0

Part IV

About the Scientific Python

Lectures

18.8. Examples for the scikit-learn chapter 680 681

 Scientific Python Lectures, Edition 2025.1rc0.dev0

• Sebastian Berg • Andreas Hilboll

• Lilian Besson • Himanshu
• Matthieu Boileau • Julian Hofer
• Joris Van den Bossche • Tim Hoffmann

19
• Michael Boyle • B. Hohl
• Matthew Brett • Tarek Hoteit
• BSGalvan • Gert-Ludwig Ingold
• Lars Buitinck • Zbigniew Jędrzejewski-Szmek
CHAPTER
• Pierre de Buyl • Thouis (Ray) Jones
• Ozan Çağlayan • jorgeprietoarranz
• Lawrence Chan • josephsalmon
• Adrien Chauve • Greg Kiar
• Robert Cimrman • kikocorreoso

About the Scientific Python Lectures

• Christophe Combelles • Vince Knight
• David Cournapeau • LFP6
• Dave • Manuel López-Ibáñez
• dogacan dugmeci • Marco Mangan
• Török Edwin • Nicola Masarone
• egens • John McLaughlin
Release: 2025.1rc0.dev0 • Andre Espaze • mhemantha
The lectures are archived on zenodo: http://dx.doi.org/10.5281/zenodo.594102 • André Espaze • michelemaroni89
All code and material is licensed under a Creative Commons Attribution 4.0 International License (CC- • Loïc Estève • K. Jarrod Millman
by) http://creativecommons.org/licenses/by/4.0/ • Corey Farwell • Mohammad

19.1 Authors • André Espaze • Tim Gates • Zachary Moon

• Emmanuelle Gouillart • Stuart Geiger • Mike Mueller
19.1.1 Editors
• Mike Müller • Olivier Georg • negm
• K. Jarrod Millman
• Fabian Pedregosa • Daniel Gerigk • John B Nelson
• Stéfan van der Walt • Robert Gieseke • nicoguaro
• Didrik Pinte
• Gaël Varoquaux • Philip Gillißen • Sergio Oller
• Nicolas Rougier
• Emmanuelle Gouillart • Ralf Gommers • Theofilos Papapanagiotou
• Gaël Varoquaux
• Olav Vahtras • Pauli Virtanen • Emmanuelle Gouillart • patniharshit
• Pierre de Buyl • Zbigniew Jędrzejewski-Szmek • Julia Gustavsen • Fabian Pedregosa
• Omar Gutiérrez • Philippe Pepiot
• Valentin Haenel (editor from 2011 to 2015)
19.1.2 Chapter authors • Matt Haberland • Tiago M. D. Pereira
Listed by alphabetical order. 19.1.3 Additional Contributions • Valentin Haenel • Nicolas Pettiaux
• Christopher Burns Listed by alphabetical order • Pierre Haessig • Didrik Pinte
• Adrian Chauve • Osayd Abdu • Bruno Hanzen • Evgeny Pogrebnyak
• Robert Cimrman • arunpersaud • Michael Hartmann • reverland
• Christophe Combelles • Ross Barnowski • Jonathan Helmus • Maximilien Riehl

682 19.1. Authors 683

 Scientific Python Lectures, Edition 2025.1rc0.dev0

• Kristian Rother • Bartosz Telenczuk

• Nicolas P. Rougier • tommyod
• Pamphile Roy • Wes Turner
• Rutzmoser • Akihiro Uchida
• Sander • Utkarsh Upadhyay Index
• João Felipe Santos • Olav Vahtras
• Mark Setchell • Stéfan van der Walt
• Helen Sherwood-Taylor • Gaël Varoquaux
• Shoeboxam • Nelle Varoquaux
• Simon • Olivier Verdier
• solarjoe • VirgileFritsch
• ssmiller • Pauli Virtanen D
• Scott Staniewicz • Yosh Wakeham diff, 567, 571
differentiation, 567
• strpeter • yasutomo57jp dsolve, 571
• surfer190
E
equations
algebraic, 569
differential, 571

I
integration, 568

M
Matrix, 570

P
Python Enhancement Proposals
PEP 255, 300
PEP 3118, 338
PEP 3129, 309
PEP 318, 302, 309
PEP 342, 300
PEP 343, 310
PEP 380, 301
PEP 380#id13, 301
PEP 8, 304

S
solve, 569

19.1. Authors 684 685