search

NVlabs / pacnet

Pixel-Adaptive Convolutional Neural Networks (CVPR '19)
https://suhangpro.github.io/pac/
Other
514 stars 79 forks source link

autocast support #27

Closed jsetty closed 2 years ago

jsetty commented 2 years ago

Hi,

I am trying to use pacnet in my model like in this example: https://github.com/NVlabs/pacnet/blob/12d52b6ebdd8e8afa0d2e54486ba77fbb3697a53/task_semanticSegmentation/fcn8s.py

However, when I try to implement Mixed-Precision training with torch.cuda.amp (with torch 1.11), I have an error in the backward method.

  File "<path>/model/pac.py", line 179, in backward
    grad_in_mul_k = torch.einsum('iomn,ojkl->ijklmn', (grad_output, weight))
  File "/opt/conda/lib/python3.8/site-packages/torch/functional.py", line 325, in einsum
    return einsum(equation, *_operands)
  File "/opt/conda/lib/python3.8/site-packages/torch/functional.py", line 327, in einsum
    return _VF.einsum(equation, operands)  # type: ignore[attr-defined]
RuntimeError: expected scalar type Half but found Float

I am using the th14 branch. Is there a fix or a work-around for this?

You may please delete this issue as this is a duplicate. Unfortunately, I cannot delete once posted.

Thanks!

suhangpro commented 2 years ago

Duplicate #24