The +0.5 term in StatQ results in asymmetric weight quantization.

[LuletterSoul]

Unlike conventional symmetric weight quantization, I found that StatQ will let rounded integer weight add a term (+ 0.5), This item may cause the StatQ not to be actually deployed on hardware.

If we remove the -0.5 and +0.5 terms, the integer value of the weight will have an extra value, that is, 2**(bit - 1) -1. How do you modify this formula and achieve the same performance?

[jiho264]

Hi, I'm not the author, but I think the +0.5 term can be decomposed like an asymmetric quantization The best explanation can be found in eq(5) in the LSQ+ paper. (LSQ+: Improving low-bit quantization through learnable offsets and better initialization) Thanks