Currently I am using export CUDA_VISIBLE_DEVICES="" to test using the only the cpu with your code.

It would be neat if we could use a --disable-cuda flag as described in pytorch device-agnostic example to use the cpu explicitely without having to mess with environment variables.

This would be useful for example when you have a GPU but not not enough VRAM to put the model on it.

Thanks, it works!

It is my understanding that by default, pytorch puts everything on the cpu and you have to specify .cuda() or .to(torch.device("cuda") to move things to the gpu.

So I think that maybe all the model.cpu() calls are not needed. Same for .to(torch.device("cpu"))

👍 Yeah you are right, just didn't think about that when I first made the PR, but I fixed it now, so it's a bit cleaner

Currently I am using export CUDA_VISIBLE_DEVICES="" to test using the only the cpu with your code.

It would be neat if we could use a --disable-cuda flag as described in pytorch device-agnostic example to use the cpu explicitely without having to mess with environment variables.

This would be useful for example when you have a GPU but not not enough VRAM to put the model on it.

Wanted to also do this, but it's pretty annoying to pass new arguments into the classes from encoder/modules.py and I wasn't really sure if I wanted to create a global variable for it or how it could best be solved, open to any ideas though