is that also an issue for GaussianMixture?

Probably since it inherits it's fit method from BaseMixture. But this fix is made in BaseMixture so it should fix both.

Probably since it inherits it's fit method from BaseMixture. But this fix is made in BaseMixture so it should fix both.

Indeed. I updated the relevant test and the changelog.

We should add the two models in the Changed models section of the what_new file.

We should add the two models in the Changed models section of the what_new file.

There is not Changed models section for 0.20.1. I think that the change log for the bug fix release is short enough that such a section is not required in this case.

Are we sure this change belongs in 0.20.1?

It's not a regression fix but I believe it's a bug fix that can fix a potentially hard to understand annoying breakage of our fit_predict contract. So I am in favor of including it in 0.20.1 (although I don't think this bug impacts that many people).

I'm ambivalent about 0.20.1. Maybe leaning towards including it?

I fixed the conflict.

The flake8 issue on travis seems like a false positive resulting from my last merge commit.

@lesteve if you ever wanted to consider a new edge case for your diff flake8 thingy ;)

@lesteve if you ever wanted to consider a new edge case for your diff flake8 thingy ;)

I can't see the false positive you are talking about, but maybe I am missing something ...

The error I see in the Travis flake8 failure:

Running flake8 on the diff in the range 1c88b3c9c..94347b3c8 (7 commit(s)):
--------------------------------------------------------------------------------
sklearn/mixture/tests/test_gaussian_mixture.py:13:1: F811 redefinition of unused 'pytest' from line 8
import pytest

Looking at the diff, it does look like you introduced a second "import pytest" (line 13) where there was already one on line 8.