scikit-learn OSX wheels, numpy compatibility

[matthew-brett]

The sklearn wheels on pypi are only compatible with (I think) numpy 1.8,
presumably because they were built with numpy 1.8. This is what
happens if you are using an earlier version of numpy, here 1.7.1:

$ python -c 'import sklearn.svm'
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/Users/mb312/.virtualenvs/sklearn-test/lib/python2.7/site-packages/sklearn/svm/__init__.py",
line 13, in <module>
    from .classes import SVC, NuSVC, SVR, NuSVR, OneClassSVM, LinearSVC
  File "/Users/mb312/.virtualenvs/sklearn-test/lib/python2.7/site-packages/sklearn/svm/classes.py",
line 1, in <module>
    from .base import BaseLibLinear, BaseSVC, BaseLibSVM
  File "/Users/mb312/.virtualenvs/sklearn-test/lib/python2.7/site-packages/sklearn/svm/base.py",
line 8, in <module>
    from . import libsvm, liblinear
  File "__init__.pxd", line 861, in init sklearn.svm.libsvm
(sklearn/svm/libsvm.c:9134)
ValueError: numpy.ufunc has the wrong size, try recompiling

The wheels built by the travis builder are compatible with numpy >=1.6.0 (Python 2) and numpy >= 1.7 (Python 3) - maybe consider using building and uploading new wheels instead of using the current wheels on pypi? Wheels get uploaded to http://wheels.scikit-image.org/

You will probably also want to apply this patch for the release, otherwise the user will get lots of spurious import warnings.

The scikit-learn-wheels repo is at https://github.com/MacPython/scikit-learn-wheels, and @ogrisel , @ChrisBarker-NOAA and I can add collaborators so y'all can trigger and update builds.

[matthew-brett]

I just added a step to apply the warnings patch, and rebuilt the wheels, so the wheels currently at http://wheels.scikit-image.org/ should work correctly. For example:

mkvirtualenv sklearn-test
pip install numpy scipy
curl -O http://wheels.scikit-image.org/scikit_learn-0.15.1-cp27-none-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.whl
pip install scikit_learn-0.15.1-cp27-none-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.whl
python -c 'import sklearn.svm'

doesn't generate warnings any more.

[hakimjonas]

Hey Matthew, I tried following your example here except with python3.4.

The wheel won't install and gives me: scikit_learn-0.15.1-cp34-cp34m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.whl is not a supported wheel on this platform. And my mac is fully updated to 10.9.4.

Any help is much appreciated.

[matthew-brett]

The wheel won't work for conda, but it should work for any other standard Python (Python.org, system python, homebrew, macports). Which python are you using?

[hakimjonas]

Actually I'm very sorry I found an error on my part. I was using the wrong pip version... the wheel does work. Thanks for your help.

[kikohs]

Hello,
I have installed the basic scientific stack with:

pip force-reinstall --upgrade numpy scipy scikit-learn
numpy: 1.8.1
scipy: 0.14.0
scikit-learn: 0.15.1
Mac OS X: 10.9.4

I have warnings on each import of sklearn.
RuntimeWarning: numpy.ufunc size changed, may indicate binary incompatibility

How can I remove the warnings or apply the patch without compiling from source?

[ogrisel]

I have just pushed @matthew-brett's scikit-learn wheels to PyPI. Can you please try to do:

pip uninstall -y scikit-learn
rm -rf ~/.pip/cache
pip install scikit-learn

[ogrisel]

The above fixes the problem reported by @kikohs on a new Pythoon 2.7 venv I just created to reproduce the problem. If you confirm that the problem is fixed on your side, we should be able to close this issue.

In the future we will try to use the automated travis based build setup developed by @matthew-brett to release the official wheels instead of manually building them on my laptop to avoid this kind of problems.

[ogrisel]

I just tested that the wheel now published for scikit-learn==0.15.1 on PyPI can work successfully with numpy 1.7.1 and Python 2.7 (this is the original bug reported by @matthew-brett).

[kikohs]

@ogrisel It works now thanks !

[GaelVaroquaux]

Good job, @ogrisel. Closing the issue.