Overflow when converting to true quantize

When converting my model from "true" quantize to "fake" quantize I realized that I had an overflow error.

One of the value of my weight tensor changed from 128 -> -128. This is due to an overflow int8 range being [-128, 127], but the scaling setting my data to the [-128, 128] range.

This error is either due to a bad computation of the scaling factor or to a missing cliping for parameters.

I see two ways to fix this issue:

1. Scale values between
   [-2^{N-1}, 2^{N-1} -1)]
   (with N being the number of bit to quantize to).
2. Scale between
   [-2^{N-1}, 2^{N-1}]
   but truncate with a Clip.

added Bug 🐛 label

changed the description

@olivierbichler @bhalimi @noamzerah

mentioned in issue #74

assigned to @noamzerah

added StatusTODO label