NVIDIA/waveglow

Python 3.8/Torch 1.6 'Conv1d' object has no attribute '_non_persistent_buffers_set'

jonaskohler opened this issue · 4 comments

jonaskohler commented

Hi there

first off, thanks for open sourcing your code. It has worked great for me in the past. I have updated to Python 3.8/Torch 1.6 lately and even after downloading your latest release

python3 inference.py -f <(ls mel_spectrograms/*.pt) -w waveglow_256channels_universal_v5.pt -o . --is_fp16 -s 0.6

gives

_torch.nn.modules.module.ModuleAttributeError: 'Conv1d' object has no attribute 'non_persistent_buffers_set'

Can you help me out please?

beckernick commented

Hey @jonaskohler . Not a maintainer of this library, but I believe this may be the following issue: https://discuss.pytorch.org/t/how-to-load-an-older-model-in-pytorch-1-6/97423

If you downgrade to 1.5.1, things might just work.

rafaelvalle commented 
for k, m in waveglow.named_modules():
   m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatability
  • 👍2
  • ❤️1
jonaskohler commented

great, thanks a lot :)!

sungjae-cho commented

It works for pytorch 1.6.0 !