quantized.pytorch Awesome work!!!But i get some questions about scale

Awesome work!!!But i get some questions about scale_fix

Open Z7zuqer opened this issue 6 years ago • 1 comments

It seems that in rangBN that scale_fix should be $\frac{1}{2*ln(n)}$ as in paper, but here scale_fix becomes scale_fix = $\frac{(0.5 * 0.35) * \sqrt{(1 + (\pi * ln(4)))} )}{(\sqrt{(2 * ln(y.size(-1)))})}$

Nov 21 '18 14:11 Z7zuqer

I also have the same doubts about how to calculate this factor

Feb 07 '23 02:02 Scarecrow1010

quantized.pytorch quantized.pytorch copied to clipboard

Awesome work!!!But i get some questions about scale_fix

quantized.pytorch
quantized.pytorch copied to clipboard