Missing function about DoReFa-Quantizer

[loveklmn]

The paper about DoReFa-Net present 3 quantization steps ( gradient, weight, and [activation || output ]). Seems that only weight was implemented. Is it a mistake? Do nni research group have any plan to do about it?
https://github.com/microsoft/nni/blob/master/docs/en_US/Compression/Quantizer.rst

[linbinskn]

Exactly, weight & output & gradient are all quantized in paper. And both weight and output quantization have been supported in NNI's DoReFa-Net currently. We think the priority of implementing gradient quantization is low since DoReFa-Net is not very suitable for deployment when comparing to another training aware quantization algorithm such as QAT.