Hi!

This project kicks ass to our https://github.com/src-d/kmcuda to major extent. I think the best way for us to move on is to integrate the missing (for source{d}) parts to Faiss. Those are:

1. Proper K-means centroid initialization instead of random sampling. This includes some high level API improvement.
2. arccos over the scalar product aka angular / proper "cosine" distance. CUDA does not notice it in terms of performance.
3. Python 3 support (well, that should be easy with Swig).

If you agree with these, I will start making PRs. If you don't, I will have to incorporate Faiss inside kmcuda as the second non-free backend. Of course, I would prefer 1.