Skip to content

Migrate WeightVector to use tempita #13358

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 7 commits into from
Closed

Migrate WeightVector to use tempita #13358

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
930e27d
[skip ci] Trigger PR
Mar 1, 2019
f40dcca
Move weith_vector to a template
Mar 1, 2019
c158a79
transform weight_vector.pxd.tp into a template
Mar 1, 2019
d1139a1
move pyx into pyx.tp
Mar 1, 2019
ef6ca46
first pass in changing doubles
Mar 1, 2019
04ce5c9
change all double that were missing
Mar 1, 2019
50de266
fix compilation
Mar 1, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion sklearn/linear_model/sgd_fast.pyx
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,7 @@ from numpy.math cimport INFINITY
cdef extern from "sgd_fast_helpers.h":
bint skl_isfinite(double) nogil

from sklearn.utils.weight_vector cimport WeightVector
from sklearn.utils.weight_vector cimport WeightVector64 as WeightVector
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This should not be here but in the template. Maybe we have some sort of clash. I need to get back to that

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

unless you also template sgd_fast, you'll have to deal with both, and switch with

if floating is float:
    do something with WeightVector64
else:
    do something with WeightVector32

Copy link
Member

@jeremiedbb jeremiedbb Mar 4, 2019 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Also this diff don't belong to this PR

from sklearn.utils.seq_dataset cimport SequentialDataset64 as SequentialDataset

np.import_array()
Expand Down
5 changes: 4 additions & 1 deletion sklearn/utils/setup.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,7 +48,10 @@ def configuration(parent_package='', top_path=None):

# generate files from a template
pyx_templates = ['sklearn/utils/seq_dataset.pyx.tp',
'sklearn/utils/seq_dataset.pxd.tp']
'sklearn/utils/seq_dataset.pxd.tp',
'sklearn/utils/weight_vector.pyx.tp',
'sklearn/utils/weight_vector.pxd.tp',
]

for pyxfiles in pyx_templates:
outfile = pyxfiles.replace('.tp', '')
Expand Down
30 changes: 0 additions & 30 deletions sklearn/utils/weight_vector.pxd
View file
Open in desktop

This file was deleted.

56 changes: 56 additions & 0 deletions sklearn/utils/weight_vector.pxd.tp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,56 @@
{{py:

"""
Efficient (dense) parameter vector implementation for linear models.

Template file for easily generate fused types consistent code using Tempita
(https://github.com/cython/cython/blob/master/Cython/Tempita/_tempita.py).

Generated file: weight_vector.pxd

Each class is duplicated for all dtypes (float and double). The keywords
between double braces are substituted in setup.py.
"""

# name, c_type
dtypes = [('64', 'double'),
('32', 'float')]

def get_dispatch(dtypes):
for name, c_type in dtypes:
yield name, c_type

}}

# cython: language_level=3

cimport numpy as np

{{for name, c_type in get_dispatch(dtypes)}}

cdef extern from "math.h":
cdef extern {{c_type}} sqrt({{c_type}} x)


cdef class WeightVector{{name}}(object):
cdef np.ndarray w
cdef np.ndarray aw
cdef {{c_type}} *w_data_ptr
cdef {{c_type}} *aw_data_ptr
cdef {{c_type}} wscale
cdef {{c_type}} average_a
cdef {{c_type}} average_b
cdef int n_features
cdef {{c_type}} sq_norm

cdef void add(self, {{c_type}} *x_data_ptr, int *x_ind_ptr,
int xnnz, {{c_type}} c) nogil
cdef void add_average(self, {{c_type}} *x_data_ptr, int *x_ind_ptr,
int xnnz, {{c_type}} c, {{c_type}} num_iter) nogil
cdef {{c_type}} dot(self, {{c_type}} *x_data_ptr, int *x_ind_ptr,
int xnnz) nogil
cdef void scale(self, {{c_type}} c) nogil
cdef void reset_wscale(self) nogil
cdef {{c_type}} norm(self) nogil

{{endfor}}
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

missing newline

108 changes: 68 additions & 40 deletions sklearn/utils/weight_vector.pyx → sklearn/utils/weight_vector.pyx.tp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,27 @@
{{py:

"""
Efficient (dense) parameter vector implementation for linear models.

Template file for easily generate fused types consistent code using Tempita
(https://github.com/cython/cython/blob/master/Cython/Tempita/_tempita.py).

Generated file: weight_vector.pxd

Each class is duplicated for all dtypes (float and double). The keywords
between double braces are substituted in setup.py.
"""

# name, c_type
dtypes = [('64', 'double'),
('32', 'float')]

def get_dispatch(dtypes):
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sorry not familiar with tempita (and the tempita doc links are broken): why do you need get_dispatch? You can't simply iterate over the dtypes list?

for name, c_type in dtypes:
yield name, c_type

}}

# cython: cdivision=True
# cython: boundscheck=False
# cython: wraparound=False
Expand All @@ -21,7 +45,9 @@ from ._cython_blas cimport _dot, _scal, _axpy
np.import_array()


cdef class WeightVector(object):
{{for name, c_type in get_dispatch(dtypes)}}

cdef class WeightVector{{name}}(object):
"""Dense vector represented by a scalar and a numpy array.

The class provides methods to ``add`` a sparse vector
Expand All @@ -31,24 +57,24 @@ cdef class WeightVector(object):

Attributes
----------
w : ndarray, dtype=double, order='C'
w : ndarray, dtype={{c_type}}, order='C'
The numpy array which backs the weight vector.
aw : ndarray, dtype=double, order='C'
aw : ndarray, dtype={{c_type}}, order='C'
The numpy array which backs the average_weight vector.
w_data_ptr : double*
w_data_ptr : {{c_type}}*
A pointer to the data of the numpy array.
wscale : double
wscale : {{c_type}}
The scale of the vector.
n_features : int
The number of features (= dimensionality of ``w``).
sq_norm : double
sq_norm : {{c_type}}
The squared norm of ``w``.
"""

def __cinit__(self,
np.ndarray[double, ndim=1, mode='c'] w,
np.ndarray[double, ndim=1, mode='c'] aw):
cdef double *wdata = <double *>w.data
np.ndarray[{{c_type}}, ndim=1, mode='c'] w,
np.ndarray[{{c_type}}, ndim=1, mode='c'] aw):
cdef {{c_type}} *wdata = <{{c_type}} *>w.data

if w.shape[0] > INT_MAX:
raise ValueError("More than %d features not supported; got %d."
Expand All @@ -57,40 +83,40 @@ cdef class WeightVector(object):
self.w_data_ptr = wdata
self.wscale = 1.0
self.n_features = w.shape[0]
self.sq_norm = _dot(<int>w.shape[0], wdata, 1, wdata, 1)
self.sq_norm = _dot(w.shape[0], wdata, 1, wdata, 1)

self.aw = aw
if self.aw is not None:
self.aw_data_ptr = <double *>aw.data
self.aw_data_ptr = <{{c_type}} *>aw.data
self.average_a = 0.0
self.average_b = 1.0

cdef void add(self, double *x_data_ptr, int *x_ind_ptr, int xnnz,
double c) nogil:
cdef void add(self, {{c_type}} *x_data_ptr, int *x_ind_ptr, int xnnz,
{{c_type}} c) nogil:
"""Scales sample x by constant c and adds it to the weight vector.

This operation updates ``sq_norm``.

Parameters
----------
x_data_ptr : double*
x_data_ptr : {{c_type}}*
The array which holds the feature values of ``x``.
x_ind_ptr : np.intc*
The array which holds the feature indices of ``x``.
xnnz : int
The number of non-zero features of ``x``.
c : double
c : {{c_type}}
The scaling constant for the example.
"""
cdef int j
cdef int idx
cdef double val
cdef double innerprod = 0.0
cdef double xsqnorm = 0.0
cdef {{c_type}} val
cdef {{c_type}} innerprod = 0.0
cdef {{c_type}} xsqnorm = 0.0

# the next two lines save a factor of 2!
cdef double wscale = self.wscale
cdef double* w_data_ptr = self.w_data_ptr
cdef {{c_type}} wscale = self.wscale
cdef {{c_type}}* w_data_ptr = self.w_data_ptr

for j in range(xnnz):
idx = x_ind_ptr[j]
Expand All @@ -104,30 +130,30 @@ cdef class WeightVector(object):
# Update the average weights according to the sparse trick defined
# here: https://research.microsoft.com/pubs/192769/tricks-2012.pdf
# by Leon Bottou
cdef void add_average(self, double *x_data_ptr, int *x_ind_ptr, int xnnz,
double c, double num_iter) nogil:
cdef void add_average(self, {{c_type}} *x_data_ptr, int *x_ind_ptr, int xnnz,
{{c_type}} c, {{c_type}} num_iter) nogil:
"""Updates the average weight vector.

Parameters
----------
x_data_ptr : double*
x_data_ptr : {{c_type}}*
The array which holds the feature values of ``x``.
x_ind_ptr : np.intc*
The array which holds the feature indices of ``x``.
xnnz : int
The number of non-zero features of ``x``.
c : double
c : {{c_type}}
The scaling constant for the example.
num_iter : double
num_iter : {{c_type}}
The total number of iterations.
"""
cdef int j
cdef int idx
cdef double val
cdef double mu = 1.0 / num_iter
cdef double average_a = self.average_a
cdef double wscale = self.wscale
cdef double* aw_data_ptr = self.aw_data_ptr
cdef {{c_type}} val
cdef {{c_type}} mu = 1.0 / num_iter
cdef {{c_type}} average_a = self.average_a
cdef {{c_type}} wscale = self.wscale
cdef {{c_type}}* aw_data_ptr = self.aw_data_ptr

for j in range(xnnz):
idx = x_ind_ptr[j]
Expand All @@ -140,13 +166,13 @@ cdef class WeightVector(object):
self.average_b /= (1.0 - mu)
self.average_a += mu * self.average_b * wscale

cdef double dot(self, double *x_data_ptr, int *x_ind_ptr,
cdef {{c_type}} dot(self, {{c_type}} *x_data_ptr, int *x_ind_ptr,
int xnnz) nogil:
"""Computes the dot product of a sample x and the weight vector.

Parameters
----------
x_data_ptr : double*
x_data_ptr : {{c_type}}*
The array which holds the feature values of ``x``.
x_ind_ptr : np.intc*
The array which holds the feature indices of ``x``.
Expand All @@ -155,20 +181,20 @@ cdef class WeightVector(object):

Returns
-------
innerprod : double
innerprod : {{c_type}}
The inner product of ``x`` and ``w``.
"""
cdef int j
cdef int idx
cdef double innerprod = 0.0
cdef double* w_data_ptr = self.w_data_ptr
cdef {{c_type}} innerprod = 0.0
cdef {{c_type}}* w_data_ptr = self.w_data_ptr
for j in range(xnnz):
idx = x_ind_ptr[j]
innerprod += w_data_ptr[idx] * x_data_ptr[j]
innerprod *= self.wscale
return innerprod

cdef void scale(self, double c) nogil:
cdef void scale(self, {{c_type}} c) nogil:
"""Scales the weight vector by a constant ``c``.

It updates ``wscale`` and ``sq_norm``. If ``wscale`` gets too
Expand All @@ -182,15 +208,17 @@ cdef class WeightVector(object):
"""Scales each coef of ``w`` by ``wscale`` and resets it to 1. """
if self.aw is not None:
_axpy(<int>self.aw.shape[0], self.average_a,
<double *>self.w.data, 1, <double *>self.aw.data, 1)
<{{c_type}} *>self.w.data, 1, <{{c_type}} *>self.aw.data, 1)
_scal(<int>self.aw.shape[0], 1.0 / self.average_b,
<double *>self.aw.data, 1)
<{{c_type}} *>self.aw.data, 1)
self.average_a = 0.0
self.average_b = 1.0

_scal(<int>self.w.shape[0], self.wscale, <double *>self.w.data, 1)
_scal(<int>self.w.shape[0], self.wscale, <{{c_type}} *>self.w.data, 1)
self.wscale = 1.0

cdef double norm(self) nogil:
cdef {{c_type}} norm(self) nogil:
"""The L2 norm of the weight vector. """
return sqrt(self.sq_norm)

{{endfor}}
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Missing newline