RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation

[miseon119]

Hi, I re-run your code in windows 11, I got this error..

if I add

torch.autograd.set_detect_anomaly(True) in your main.py before I train. I got this .

How should I fix this problem.

my environment is:

• cuda 11.6
• rtx 3090
• anaconda prompt
• python 3.8

[miseon119]

I think it is a torch version problem. I used pytorch 1.12. My colleague changed code like this, it works:

[lizhe00]

Maybe such a modification cannot work, since no gradient is propagated from the total loss.

[miseon119]

You are right. It's not work.