@jnothman kmeans doctest fails because the cluster centroid changed, following is the depiction of doctest clusters before and after the fix.

should I update the expected cluster result in doctest per the new centroid or try different test data?

Strange that this doctest result changes... the doctest hasn't been changed in 40e6c43 or subsequently :|

But it's only failing on the latest numpy and scipy... so it's probably due to an upstream change. Can you change the example to make it deterministic? For example, move the top-right point further right?

I can reproduce the issue and I confirm that this PR fixes it.

I have updated the doctest to be more deterministic by moving a set of cluster points far apart from the other.

Please add an entry to the change log for version 0.20.3 at doc/whats_new/v0.20.rst. Like the other entries there, please reference this pull request with :issue: and credit yourself (and other contributors if applicable) with :user:

And seems that we have a similar problem

Btw, why do current code use all the cores when n_clusters=1 (and not when n_clusters>1)?

I can't reproduce that. However, it might be that the multithreading comes from MKL in that case. Running the snippet with the env variable MKL_NUM_THREADS=1 makes things easier to check. In that case only one core is used in any situation (before this PR). And all cores are used in any situation after this PR.

I can't reproduce that. However, it might be that the multithreading comes from MKL in that case. Running the snippet with the env variable MKL_NUM_THREADS=1 makes things easier to check. In that case only one core is used in any situation (before this PR). And all cores are used in any situation after this PR.

Thanks, I don't have time to reproduce but your version seems more reasonable. According to the code, before this PR, there's no parallelism at scikit-learn level. I guess the reporter make some mistakes accidentally.

And seems that we have a similar problem
scikit-learn/sklearn/decomposition/dict_learning.py

Line 303 in ff46f6e

if effective_n_jobs(n_jobs) or algorithm == 'threshold':

Feel free to fix it in another PR, otherwise we'll open an issue.

Sure will fix it in another PR

Thanks @nixphix