Thanks for the fix.
Can you please add a non-regression test?

I didn't check in detail but couldn't there also be a problem if centers is fortran ordered?

We could also make sure that X is C contiguous, have you done any benchmarks by any chance?

We could also make sure that X is C contiguous, have you done any benchmarks by any chance?

Yes, when I use PCA.

X = np.array([[0, 0], [0, 1]])
pca = PCA()
new_X = pca.fit_transform(X)
X.strides
>>> (4, 8)
new_X.strides
>>> (8, 4)

I didn't check in detail but couldn't there also be a problem if centers is fortran ordered?

I think you are right. It also be a problem.

Repository health increased by 0.01% when pulling a6cf751 on zhaipro:bug_k_means into 6560a8e on scikit-learn:master.

• 4 new problems were found (including 0 errors and 1 code smell).
• 5 problems were fixed (including 0 errors and 5 code smells).

Indeed this is bad bug. Could you please add a non-regression test on using toy fortran-aligned data by using np.asfortranarray?

I will perfect it as soon as possible.

@ogrisel the thing is that we should probably add a common test for this, right?

+1 for a common test.

Coverage increased (+0.0%) to 95.15% when pulling e77d36e on zhaipro:bug_k_means into 6560a8e on scikit-learn:master.

Can you please fix the random_state argument of KMeans in your test, e.g. random_state=42. That should not change anything on such a small dataset but it better to always fix the RNG in tests to make the failure reproducible and independent of other tests.

@zhaipro have you check that your new test fails when run against the code in the current master?

Could you also please squash your intermediate commits together, ideally keeping one for the FIX and one for the TST.

have you check that you test fails when run against the code in the current master?

@ogrisel Yes, I checked.

just force-push the squashed commits into your bug_k_means branch.

Fixed in #4531.