Optimize Poincare model training

[koiizukag]

Reduce computational complexity O(N^2) => O(N).
This part is the bottleneck of the overall training phase.

[piskvorky]

Great catch @koiizukag .

How did you find this? Any more examples of weak programming we could fix, while at it?

[koiizukag]

Thanks @piskvorky .

I found this insufficient implementation via python profiler.

• https://github.com/rkern/line_profiler
• https://github.com/what-studio/profiling

If we use multiprocessing in independent for loop, we can make this model faster.
ex: https://github.com/RaRe-Technologies/gensim/blob/develop/gensim/models/poincare.py#L545
Also, it is better to use Cython like other gensim models.

[mpenkov]

Thank you for your contribution @koiizukag.

@piskvorky This looks good to merge to me. Is there anything left on this PR that you want to do before we merge this?

[piskvorky]

LGTM. My only question was around fixing other similar issues in Poincare, since we (@koiizukag ) are at it. But nothing blocking.