extract_f0_print.py torch.nn.utils.weight_norm deprecation warnings

[alexlnkp]

Warnings

torch/nn/utils/weight_norm.py:28: UserWarning: torch.nn.utils.weight_norm is deprecated in favor of torch.nn.utils.parametrizations.weight_norm.

Are being spewed out during feature extraction
Fix for that should be relatively easy, just replacing torch.nn.utils.weight_norm to torch.nn.utils.parametrizations.weight_norm

[fumiama]

I will do this when I refactor training.

[TheTrustedComputer]

Removing the deprecation warning is understandable to stay current, but I have to rely on an older build of PyTorch (2.0.0) with ROCm acceleration (5.2) to enable local training on my AMD RX 5000 series GPU. This is the latest official build without resorting to compilation from source.

It's a breaking change I'm certain nobody wants in their environment, but I have a Docker container that holds my custom build of the newest PyTorch and ROCm 5.4 to keep up with development. This removal requires PyTorch 2.1+, which doesn't have wheels with the ROCm version I need.

[fumiama]

Thanks for your mention. I will add a try-except fallback.

[alexlnkp]

Removing the deprecation warning is understandable to stay current, but I have to rely on an older build of PyTorch (2.0.0) with ROCm acceleration (5.2) to enable local training on my AMD RX 5000 series GPU. This is the latest official build without resorting to compilation from source.

It's a breaking change I'm certain nobody wants in their environment, but I have a Docker container that holds my custom build of the newest PyTorch and ROCm 5.4 to keep up with development. This removal requires PyTorch 2.1+, which doesn't have wheels with the ROCm version I need.

Hmm... I guess we should stick to it for now, the plans to remove the torch.nn.utils.weight_norm from torch are not that immediate.
For now, we should use the workaround to keep up on the latest torch but we'll leave the original torch.nn.utils.weight_norm in place until torch has been updated for ROCm.

Please keep us informed on the matter.

[TheTrustedComputer]

Hmm... I guess we should stick to it for now, the plans to remove the torch.nn.utils.weight_norm from torch are not that immediate. For now, we should use the workaround to keep up on the latest torch but we'll leave the original torch.nn.utils.weight_norm in place until torch has been updated for ROCm.

Please keep us informed on the matter.

I'm not sure what you mean by "until torch has been updated for ROCm"? Well, my custom build with the newest PyTorch and ROCm 5.4 runs perfectly fine with my card, even without the HSA_OVERRIDE_GFX_VERSION override. Unfortunately, ROCm 5.2 no longer compiles against the most recent PyTorch and has known bugs I've encountered that have been fixed upstream in later releases, so I need to bump it to version 5.4, which I've heard works flawlessly with the 5500 XT; I have two of each for training purposes.

I'd say use torch.nn.utils.parametrizations.weight_norm for PyTorch >= 2.1 and torch.nn.utils.weight_norm for PyTorch <= 2.0. Those are my two cents to support the broader user base of different PyTorch versions tailored to their needs.

[fumiama]

In this case, the try-except fallback can be a good trade-off.