[MRG] MAINT: pin numpy version in windows

[glemaitre]

Trying to lower the NumPy version when building the wheels.

[ogrisel]

The failure on Python 3.5 is weird. Maybe you can try with numpy 1.12.1 for Python 3.5 so see it this is a bug in numpy?

[jorisvandenbossche]

If that is the case, I would still build with numpy 1.10, but test with numpy 1.12. In the travis configuration a similar distinction is used.

[ogrisel]

If that is the case, I would still build with numpy 1.10, but test with numpy 1.12. In the travis configuration a similar distinction is used.

I agree.

[glemaitre]

I am doing a build bumping numpy. I will try to investigate the failure in a virtual machine.

[ogrisel]

Thanks!

[glemaitre]

So the reason is that np.clip does not discard NaN in this version

In [1]: import numpy as np

In [2]: X = np.array([np.nan, 1, 2, np.nan, 10])

In [3]: np.clip(X, 0, 5)
Out[3]: array([ 0.,  1.,  2.,  0.,  5.])

in the other version, we get something like:

array([np.nan, 1., 2., np.nan, 5.])

I'll try to check why.

[ogrisel]

As @jorisvandenbossche said, build with the wheel oldest version of numpy and then run the tests with a newer version of numpy that does not have the bug. No need to debug old versions of numpy.

[glemaitre]

I am testing with an older version. The bug is actually this one:
numpy/numpy#7601

The issue should be solved in 1.11.3 wheel.

[ogrisel]

I think we should always run the tests with the newest stable version of numpy at the time of the newest scikit-learn release. WDYT @jorisvandenbossche @glemaitre ?

[glemaitre]

I don't have strong opinion but it could be the most common case for most users.

[ogrisel]

Thanks for the fix!