Closed alex-ong closed 4 years ago
Since we use it heavily in text recognition, we should benchmark it. numpy.diff, numpy.abs and numpy.sum are the key functions; which probably don't need BLAS
numpy.diff, numpy.abs and numpy.sum
Every other usage of numpy is just for creating numpy arrays.
Based on this specification, we don't need BLAS: http://www.netlib.org/blas/#_level_1
Wheels for windows: https://www.lfd.uci.edu/~gohlke/pythonlibs/#numpy
I will post benchmark soon
Running digit.py: 1.18.5 with OpenBLAS -> 10000 iterations took:10.113064289093018 seconds 1.16.6+vanilla -> 10000 iterations took:9.158168315887451 seconds
1.18.5 with OpenBLAS -> 10000 iterations took:10.113064289093018 seconds
1.16.6+vanilla -> 10000 iterations took:9.158168315887451 seconds
So we can use vanilla for now.
Since we use it heavily in text recognition, we should benchmark it.
numpy.diff, numpy.abs and numpy.sum
are the key functions; which probably don't need BLASEvery other usage of numpy is just for creating numpy arrays.
Based on this specification, we don't need BLAS: http://www.netlib.org/blas/#_level_1
Wheels for windows: https://www.lfd.uci.edu/~gohlke/pythonlibs/#numpy
I will post benchmark soon