TST Ignore Kmeans test failures on MacOS #12648
Conversation
|
I would rather submit this just to the 0.20.X branch and in the mean time try to actually fix the issue on master. |
|
Maybe we'll add MacOS CI in master before the fix and I guess it's not harmful to include it in master? |
|
why skip instead of v-measure? But I don't have a strong opinion. |
|
We'll eventually fix the issue so I guess there's not so much difference. I use pytest.xfail because we don't need to rewrite the test after we fix it (I guess in this case, we should get identical labels). |
|
I agree with @qinhanmin2014, fit_predict and fit.predict should give the same labels, not a permutation. |
|
Right now I'm mostly concerned about the failure on conda-forge: conda-forge/scikit-learn-feedstock#81 |
|
Untagging from 0.20.1 since this same PR was merged to 0.20.X directly in #12651 as suggested by Olivier That leaves more time to come up with a proper solution for it on master.. |
|
Seems that the failure disappears. Let's reopen if the error persists. |
|
#13751 is a nicer fix, imho
|
|
Please note my comment in the code, @qinhanmin2014, and merge if happy. |
| # which appears to be where it fails on some MacOS setups. | ||
| # NB: This test is largely redundant with respect to test_predict and | ||
| # test_predict_equal_labels. This test has the added effect of | ||
| # testing idempotence of the fittng procesdure which appears to |
There was a problem hiding this comment.
typo: "fittng procesdure" but fair enough :)
There was a problem hiding this comment.
I'm still not used to the keyboard on these new macbooks
|
@jeremiedbb can we revert this xfailed test with the new KMeans implementation? |
|
I'd say probably, but there's still a possibility of failure because with elkan, the predict method still uses lloyd algo whereas in fit_predict it's elkan all the way. I'm working on a refactoring of kmeans which will fix that. So I think we can keep the skip a liitle more. |
See #12644 (comment)
Tag #12644 from 0.20.1 to 0.21