search

alex-ong / NESTrisOCR

OCR for statistics in NESTris
24 stars 7 forks source link

Investigate ocr_algo.board numba performance #49

Closed alex-ong closed 4 years ago

alex-ong commented 4 years ago

Doing some quick tests on pre-compiling vs JIT.

I noticed we should use at least a few sample boards, otherwise JIT will give wrong result as it will over-fit.

alex-ong commented 4 years ago

I hope the njit isn't overfitting its optimizations, since we only have one test case.

2000 tests / alternating 2 images

numba njit njit \ nogil \ cache precompile
0.47.0 0.000168764 0.000074922 0.000256539
0.49.1 0.000176369 0.000084009 0.000240302

tl;dr: for dev, lets use njit | nogil | cache.

For distribution, we'll still use precompilation, because packaging numba is +50MB