torch_device errors in latest builds about torch

[dfalbel]

Thanks for the investigation @shikokuchuo ! This was really helpful! I think #1122 will fix it

[shikokuchuo]

Breaking this down after further investigation:

The "y is not a torch_device" error is actually to do with the use of safetensors. If I set torch.serialization_version to 2 then things work as before. It points to this error occurring during loading of serialised tensors (model and optimiser) saved previously. I haven't had the time to figure out exactly where it happens, but it seems that this is a breaking change for existing code.

The second "length" error only happens within Rstudio, I'm guessing as it automatically tries to determine it's length for the environment pane or something. More of a cosmetic issue.

[shikokuchuo]

Thanks I see you've addressed the second point already.

I've pinpointed the first issue. It seems that the 'device' argument of safetensors::safe_load() only takes a character string and not a torch device already created with torch_device().

Specifically where this was causing an issue was with loading of an optimizer state dict to the correct device, in the context of something like the code below:

device <- torch_device("cuda:0")

net <- torch_load(modelfile)
net$to(device = device)$train()
optimiser <- optim_adam(net$parameters)
optimiser$load_state_dict(torch_load(optimfile, device = device))

This worked before, but now the second torch_load() fails, but works if torch_load(optimfile, device = "cuda:0") is specified instead.

[shikokuchuo]

Thanks, can confirm this fixes the issue.