fangwei123456 / pytorch_entropy_loss Goto Github PK

Hi,
Thank you for your implementation of calculating the entropy loss using pytorch. It is really useful because (I think) all the DL model for image compression task used some kind of techniques to approximate the entropy loss in the training stage.
I tried to run this repo on my local enviroment, and I found that the torch.histc does not support a cuda tensor as its input. Therefore in the entropy_loss.py:

p[i] = torch.histc(input[i].view(-1).int(), bins = maxV - minV + 1, min = minV, max = maxV).float() / input[i].numel()

it seems not working.

I think it might be changed to:

p[i] = torch.histc(input[i].view(-1).int().data.cpu(), bins = maxV - minV + 1, min = minV, max = maxV).float() / input[i].numel()

where we might need to put input[i] on the "CPU memory".

Also, I guess that you are working on the reproduction of the algorithm mentioned in paper VARIATIONAL IMAGE COMPRESSION WITH A SCALE HYPERPRIOR. In the quantification stage, you might need to add a uniform noise as it is mentioned in the paper.

I'm also working on the reproduction of this paper, I wonder if you could share your progress? Because I faced some issue when implementing the hyperprior module and also my "factorized prior" version cannot match the performance mentioned in the paper.

Looking forward for your reply,

Gong