Thanks for the report. Does any-one know how/whether pip checks dependencies? And where the error comes from?

I think we need a requirements setting in setup.py
Check out nltk's setup vs ours.
https://github.com/nltk/nltk/blob/master/setup.py#L84

https://github.com/scikit-learn/scikit-learn/blob/master/setup.py#L111

On 02/17/2013 04:51 PM, Hrishikesh Huilgolkar wrote:

I installed pip install scikit-learn without scipy. It did not give me
any error.
When i imported sklearn in ipython, it game me error, could not find
scipy.
So this issue is resolved?

—
Reply to this email directly or view it on GitHub
#1495 (comment).

I guess it is running the tests that results in the error.

Adding an install_requires doesn't work, probably because we use numpy.distutils.core.setup instead of setuptools.

Numpy itself requires an older version of itself or else errors during installation.

numpy/numpy#2990

Resolved by installing

pip install scipy

I avoided using specific version, earlier I was using latest version of scipy that might be the reason I was getting error.

+1 for at least improving the error message triggered by sklearn/__init__.py.

Wouldn't the "right" way be to add scipy to the requirements.txt. I guess we don't do that because we don't want to make people pip install numpy? Also see #4164

@larsmans we could use build-requires as scipy does, right:
https://github.com/scipy/scipy/blob/master/setup.py#L190

Wouldn't the "right" way be to add scipy to the requirements.txt.

It would probably lead to some horrible message crashing because of the
lack of a fortran compiler. I'd much rather have people that do not have
scipy installed go and read the install instructions (ours, or scipy's)
to see that it is strongly advised not to install scipy from source.
Relying on the automatic dependency crawling actually only works for
advanced users who has a sane build environment, and paradoxically not
for Joe users who think that they might benefit from it.

How about in setup.py we do a

try:
    import numpy
except ImportError:
   raise SomeError("Install a decent pre-compiled numpy")
try:
   import scipy
except ImportError:
   raise SomeError("install a decent scipy")

How about in setup.py we do a

That sounds good (with a good error message, probably pointing to online
docs with a summary of the discussion that we just had).

Exactly.

Possibly even without the error traceback would be better, i.e. print to
stderr and exit -1.

On 27 January 2015 at 10:00, Andreas Mueller notifications@github.com
wrote:

Exactly.

—
Reply to this email directly or view it on GitHub
#1495 (comment)
.

Possibly even without the error traceback would be better, i.e. print to
stderr and exit -1.

No: errors are mean exactly for this: to be captured elsewhere. An
expection is a machine understandable error message, unlike something
printed to stderr.

Mhmmm. I guess it is actually an ImportError and I wasn't thinking of the
case where the first use is through some other software, which is, of
course, common. I stand corrected.

On 27 January 2015 at 10:59, Gael Varoquaux notifications@github.com
wrote:

Possibly even without the error traceback would be better, i.e. print to
stderr and exit -1.

No: errors are mean exactly for this: to be captured elsewhere. An
expection is a machine understandable error message, unlike something
printed to stderr.

—
Reply to this email directly or view it on GitHub
#1495 (comment)
.