2

For various hardware and software performance reasons I want to force a symmetric quantization in TensorFlow. I figure the quickest way to try this out and compare its accuracy with the existing asymmetric quantization is to set the quantization max to be the negative of the min everywhere in "ls tensorflow/core/kernels/quant". Am I missing any other files that need to be modified? Is there a better more centralized way to achieve the desired behavior?

Sean
  • 21
  • 1

1 Answers1

0

This commit added SCALED mode, which may be what you want.

With the traditional MIN_COMBINED mode, the zero point is shifted by 0.5 when you "set the quantization max to be the negative of the min". That's because -|max| maps to -128 and |max| maps to 127.

Jingyue Wu
  • 161
  • 2