Ristretto based Caffe_DFP: all caffemodel conv layer weights are zero

[Yixin Du]

Dear all,

I'm trying to train a simple residual learning based loop filter (LF) using Caffe_DFP, which is a modified version of Ristretto Caffe. However, after training, I found that all the weights of conv layers in caffemodel are zero. While the training loss did decrease from 1000 to 0.3. This zero weights problem also happens when I tried to train another simple network such as VDSR using Caffe_DFP. As a comparison, when I train LF (without quantization) or VDSR using official caffe, this issue does not exist.

The repository for Caffe_DFP can be found at:  https://github.com/Hikvision-Codec/Caffe_DFP

I have attached my train.prototex and solver, as well as the loss plot. I'd appreciate if someone could help with this. 

Thank you very much!

Yixin