[Quantization] Calculate Scale and Offset based on the histogram

[rdzhabarov]

Currently, we profile all nodes in a graph and capture both min/max as well as the histogram.

Quantization scale and offset are calculated based on the global min and global max and do not take histogram into consideration.
This does not work very well when there are outliers.

There was the logic for calculating scale and offset based on histogram but it was removed: b5d9385
As it did not properly work. Ex, simple tensor with {1, 2} values.

[jspark1105]

Let me take a look if the l2 minimization quantization can also work robustly even for very small tensor like {1,2}.

[jspark1105]

I'm seeing l2 minimization quantization gives you scale 2/255 and offset 0 (as if min was 0 and max was 2). What results did you see?

[rdzhabarov]

@jspark1105 just to clarify, do you supply histogram of 1 (corresponds to 1) 000000...01 (corresponds to 2) to l2 minimization? Your results look good.

[jspark1105]

Yes.

[rdzhabarov]

hm, thanks for checking. I'll take a closer look at this.