LookSAM bug #442
Description
I just want to report I bug I meet using LookSAM. It is at
sam.py", line 818, in second_step
grad, grad_norm * torch.sum(g_grad_norm * g_s_grad_norm) * g_grad_norm
~~~~~~~~~~~~^~~~~~~~~~~~~~~
RuntimeError: The size of tensor a (16) must match the size of tensor b (3) at non-singleton dimension 3
It happens when doing .second_step() on the optimizer.
Unfortunately I am not able to provide a code to reproduce the error since it came after custom modules and loss built inside a custom pytorch wrapper used to build network architectures with pytorch list. I just wanted to report this. The same code choosing SAM, FriendlySAM, WSAM and GSAM worked properly.
- OS : Windows
- PyTorch version : 2.8
- Python version : 3.12
- pytorch-optimizer version : the one present on github at the present moment (with pip I got an older one with less functions)