If I understand correctly, in the current setup appveyor installs the scipy/numpy wheels from http://www.lfd.uci.edu/~gohlke/pythonlibs/ which are built against MKL,

scikit-learn/build_tools/appveyor/requirements.txt

Lines 1 to 11 in 4d9a12d

    
           # Fetch numpy and scipy wheels from the sklearn rackspace wheelhouse. 
        
           # Those wheels were collected from http://www.lfd.uci.edu/~gohlke/pythonlibs/ 
        
           # This is a temporary solution. As soon as numpy and scipy provide official 
        
           # wheel for windows we ca delete this --find-links line. 
        
           --find-links http://28daf2247a33ed269873-7b1aad3fab3cc330e1fd9d109892382a.r6.cf2.rackcdn.com/ 
        
           # fix the versions of numpy to force the use of numpy and scipy to use the whl 
        
           # of the rackspace folder instead of trying to install from more recent 
        
           # source tarball published on PyPI 
        
           numpy==1.9.3 
        
           scipy==0.16.0

so that would make scikit-learn wheels on windows also dependent on MKL?

The latest versions of numpy / scipy do have windows wheels on PyPi (according to what you said, built with OpenBlas). However, that's not the case for the versions currently pinned in appveyor. I'm not sure how backward compatibility works here: would it make sense incrementing the numpy/scipy versions in appveyor for v0.20 to those that have an official windows wheel, or not?

cc @ogrisel

Uh oh!

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions