Remove _hmmc.pyx; rewrite functions from this file as python+numpy #1825
Conversation
…nctions. On my tests, this makes .fit (with 100 states) 3x faster, .eval (with 1000 states) 20% faster, .decode_viterbi (with 1000 states) 7x faster.
|
Wow. +1. |
|
It would be nice to see how much could be gained by replacing the ndarray in the function definition by typed memory views. |
|
Very nice indeed! Thanks a lot.
Yup, but I don't think that this suggestion should delay merging this |
|
I'll take a look at possible speed gains in Cython extension in a couple of hours. |
|
_forward, _backward and _compute_lneta had Python calls because _logsum was not cdef'ed (typed memoryviews has nothing to do with this - they could make code faster, but I didn't try them now because the main issue was not with ndarray). This version: kmike@ab8b8a2 is 10-20% faster than in this pull request regarding .fit() and .eval() with 100 and 1000 states; with 2 states .fit and .eval are 5x faster with Cython than with vectorized numpy; _viterbi is still slow there (we could bring it on par just by copying python+numpy version). |
|
With #1828 merged, this one can be closed, right? Feel free to reopen if I'm wrong. |
On my tests, this makes .fit (with 100 states) 3x faster, .eval (with 1000 states) 20% faster, .decode_viterbi (with 1000 states) 7x faster.
It might be possible to write a better Cython extension, so I didn't inline functions extracted from _hmmc.pyx to _BaseHMM methods.