jorge-pessoa / pytorch-gdn Goto Github PK

I get many of these errors
/build/python-pytorch/src/pytorch-1.3.1-opt-cuda/torch/csrc/autograd/python_function.cpp:620: UserWarning: Legacy autograd function with non-static forward method is deprecated and will be removed in 1.3. Please use new-style autograd function with static forward method. (Example: https://pytorch.org/docs/stable/autograd.html#torch.autograd.Function)

Likely from the LowerBound(Function) class. Luckily it still works in pytorch -1.3.1

I've encountered an interesting issue: when the number of channels in GDN exceeds 181, the CPU usage becomes remarkably high. It seems like 181 is a critical threshold, but I'm not sure if others have experienced the same problem

What is the point of your LowerBound custom autograd function? Is the functionality different than standard torch.max(x, b) or F.threshold(x,b,b)? I don't see a situation where your variable pass_through_2 in .backward() would ever be true (unless you have a negative reparam_offset).

Hi, thanks for sharing! I found a little mistake in your comment of GDN's math equation. In my opinion, the x has been squared.

In your code:

IMO this equation should be

y[i] = x[i] / sqrt(beta[i] + sum_j(gamma[j, i] * x[j]^2))

Please check for this.

When trying to export a pytorch model containing GDN to ONNX, it complains about:
RuntimeError: ONNX export failed: Couldn't export Python operator LowerBound

Has anyone tried this? I don't know for the moment whether we have to add the LowerBound 'operator' to onnx, or
rewrite GDN in some way to avoid the 'operator'.

Hi,
Thank you for this implementation.
This GDN module works well on a single GPU, but when I warp my model into the Data parallel module and use multi-GPUs to train my model, the training failed. Could you please check it?