Minimum requirements for dependecy for Windows wheels

I got the CI from imbalanced-learn screaming at me after installing the RC. The reason is that the wheel is built with numpy 1.14.5 and conda installed numpy 1.13.3. Since the requirements are 1.8.2, there is no upgrade when installing scikit-learn.

@jnothman @amueller What are you thoughts?

Step that should reproduce the error:

conda create --name sklearn-test python=3.5
conda install numpy scipy -y -q
pip install --pre scikit-learn

python -c "import sklearn"

pytest --pyargs imblearn --cov-report term-missing --cov=imblearn
============================= test session starts =============================
platform win32 -- Python 3.5.2, pytest-3.7.3, py-1.5.4, pluggy-0.7.1
rootdir: C:\projects\imbalanced-learn, inifile: setup.cfg
plugins: cov-2.5.1
collected 0 items / 123 errors
=================================== ERRORS ====================================
____________________ ERROR collecting imblearn/__init__.py ____________________
__init__.pxd:998: in numpy.import_array
    ???
E   RuntimeError: module compiled against API version 0xc but this version of numpy is 0xb
During handling of the above exception, another exception occurred:
..\imblearn\__init__.py:35: in <module>
    from .base import FunctionSampler
..\imblearn\base.py:14: in <module>
    from sklearn.base import BaseEstimator
C:\Miniconda35-x64\lib\site-packages\sklearn\__init__.py:64: in <module>
    from .base import clone
C:\Miniconda35-x64\lib\site-packages\sklearn\base.py:13: in <module>
    from .utils.fixes import signature
C:\Miniconda35-x64\lib\site-packages\sklearn\utils\__init__.py:12: in <module>
    from .murmurhash import murmurhash3_32
sklearn\utils\murmurhash.pyx:26: in init sklearn.utils.murmurhash
    ???
__init__.pxd:1000: in numpy.import_array
    ???
E   ImportError: numpy.core.multiarray failed to import

[rth]

Despite our requirements above it’s not possible to build scikit-learn on Python 3.7 with numpy 1.14.0 because it is unsupported by numpy. I would argue the same applies to all other Python versions (Py3.6, Py3.5) with the minimal versions of numpy/scipy to be determined (and that are unrelated to our minimum requirements).

On top of that, Windows has some additional complications, which explain for instance why in conda forge the minimum numpy requirement is higher on Windows than on Linux.

I don’t understand all the details, just saying that determining the minimum numpy/scipy versions for each Python version and OS is non trivial (or at least time consuming) and this is not necessarily related to the minimal numpy/scipy versions we have in the docs.

[ogrisel]

I merged MacPython/scikit-learn-wheels#11 . Sorry for the original mess, I think I was sloppy with the numpy build version to workaround a failing test on old numpy in some entry of the build matrix. The new setup with distinct versions of numpy for building and testing is better.