[MRG] Add Windows continuous integration with AppVeyor CI #3363
Conversation
|
I forgot to link to a working sample output of this build setup. Here is the report I get when I configured AppVeyor to watch my own fork of the sklearn github repo: https://ci.appveyor.com/project/ogrisel/scikit-learn Note that you can click on the artifact tab of each sub-entry of the build matrix to download the generated |
|
Once this is merge in master, I will reconfigure https://ci.appveyor.com/project/ogrisel/scikit-learn to point to https://github.com/scikit-learn/scikit-learn . Then existing PRs will have to be rebased on top of master to get properly tested under windows. Ideally I would also like to backport that to 0.15.X to be able to use AppVeyor to build the binary release artifacts for Windows for scikit-learn 0.15.0. |
| @@ -603,3 +604,18 @@ def check_skip_network(): | |||
|
|
|||
|
|
|||
| with_network = with_setup(check_skip_network) | |||
|
|
|||
|
|
|||
| def is_32bit(): | |||
There was a problem hiding this comment.
I would add this to the __ALL__constant and add a small docstring.
|
@arjoly I addressed your comment and decided to disable the 32 bit windows skip to see exactly which estimators are failing. I just tried to run the tests on Python 3.4 32-bit on a Rackspace windows box and could not reproduce the failures any more in 10 consecutive runs. Let see if appveyor is failing or not. |
|
Hm, there is one failed build. |
|
Yes I am fixing a bunch of other heisen tests before tackling #3255. |
|
I rebased on top of the current master where I pushed some fixes for other tests and rewrote this PR to only skip the |
| # To avoid spurious builds and notification as long as not merged in master | ||
| branches: | ||
| only: | ||
| - appveyor-ci |
There was a problem hiding this comment.
How about we remove this for the merge!
|
👍 to merge with these 2 comments, if the tests pass. |
| @@ -0,0 +1,6 @@ | |||
| --find-links http://28daf2247a33ed269873-7b1aad3fab3cc330e1fd9d109892382a.r6.cf2.rackcdn.com/index.html | |||
There was a problem hiding this comment.
It's a temporary wheelhouse we host on the rackspace cloud account of the sklearn project to host numpy and scipy wheels built from Christoph Gohlke's installers. I will add an inline comment.
There was a problem hiding this comment.
s/we ca delete/we can delete
Thanks for the clarification. The note about how they're obtained is definitely useful.
|
Changes Unknown when pulling 1481e53 on ogrisel:appveyor-ci into * on scikit-learn:master*. |
|
I rebased on master and it's all green (both appveyor and travis, although travis should not be affected). |
|
OK, merging this guy. |
[MRG] Add Windows continuous integration with AppVeyor CI
AppVeyor is a system similar to Travis CI that provides a Windows based build environment with MSVC++ Express 2008 and 2010 pre-installed along with the matching Windows SDKs for 64 bit builds. This PR configures how to build scikit-learn, generate .whl and .exe packages for 4 supported target platforms: Python 3.4 & 2.7, each with 32 and 64 bit.
This PR also has a workaround for #3255: some common tests that call
fit_transformtwice (with a fixed random state) do not deterministic output on 32 bit Python (both 2 and 3) for some estimators that use LAPACK routines like SVD or eigen solvers. Note the numpy / scipy libraries tested here come from Christoph Gohlke's distribution and are therefore linked with MKL. However I also had non-deterministicfit_transformfor the same bunch of estimators when I manually built numpy and scipy with OpenBLAS on a Rackspace windows VM a while ago so the problem is not tied to the use of MKL. I therefore decided to catchAssertionErrorand raiseSkipTestinstead for this specifictest_commoncheck under 32 bit platforms.