I am trying to use mkldnn to speed up training. My model is fasttext. I have followed the instruction of pytorch document (from source). But it seems that sometimes the training is still too slow. And also I want to know how many benefits I can gain by using mkldnn.
Try this:
import torch
print(*torch.__config__.show().split("\n"), sep="\n")
You will get the build information, if USE_MKLDNN=ON
then you are good to go.
PS: this works for PyTorch 1.1 or earlier.