Issue with import from LatentDirichletAllocation

Some people are getting an issue after updating scikit-learn:

  File "/home/jupyter/.local/lib/python3.5/site-packages/sklearn/decomposition/__init__.py", line 19, in <module>
    from ._online_lda import LatentDirichletAllocation
ImportError: cannot import name 'LatentDirichletAllocation'

There an SO question about it: https://stackoverflow.com/questions/59240556/latentdirichletallocation-was-not-installed-in-decomposition-module-of-scikit-le

and we got an issue reported in imbalanced-learn: https://github.com/scikit-learn-contrib/imbalanced-learn/issues/663

I assume that this is not just bad luck. However, I cannot reproduce it locally. It is for sure linked to the privatization of the online_lda module but I don’t get how to trigger that.

NB: I will write an answer in the so because the “solution” is a really bad fix and if we could get feedback on this issue to solve it, it would be great.

[glemaitre]

OK so the issue is due to the following file _online_lda.cpython-36m-x86_64-linux-gnu.so

Before we had a _online_lda.pyx and therefore we were generated the associated so file. We renamed it to _online_lda_fast.pyx. In the same time, we renamed online_lda.py to _online_lda.py.

Basically, during pip install scikit-learn -U, the .so does not get removed and we have a collision between the naming. To solve this issue, you need to remove the file by hand.

[NicolasHug]

Did we have some other cases where we renamed the pyx file and took the same name for the py file. We will have the same issue for this one as well.

I think _k_means.py[x] and _hierarchical.py[x] also qualify