RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn about cca_zoo HOT 4 CLOSED

From a quick google of this error (I can't replicate without perhaps more info on versions of pytorch?) it looks like it might relate to mixing floats and doubles.

Depending on your use case take a look at the pytorch-lightning branch where I have written a new training loop wrapper using pytorch-lightning which leverages all of their behind the scenes stuff with device and data management and is a much preferable option to my inflexible wrapper!

from cca_zoo.

Hi, I work with PyTorch version 1.8.0, it's very strange that it works for you and I keep getting this bag. For some reason, it has a problem with the gradient calculation.

I found the problem, however, I don't know how to fix it.
The problem is in line 217 in the file: deepmodels/objective.py, where you calculate:
eigvals = torch.linalg.eigvalsh(trace_TT)
This command changes the grad_fn to None and arises the problem.

from cca_zoo.

After finding the bug reason, and reading the following issue:pytorch/pytorch#57189, the problem was solved with the update of PyTorch to a newer version.
Thank you for your help!

from cca_zoo.

Ah thanks for hunting this down! Will reopen this as a Reminder to myself to change the requirements to specify this version of torch.

from cca_zoo.