Use of hooks about pytorch-cnn-visualizations HOT 3 CLOSED

Hello, unfortunately, I don't have a good answer to you other than: autograd just expects a tuple. You can change the tuple to a tensor (or any other thing) and get the following error:

Traceback (most recent call last):
  File "guided_backprop.py", line 74, in <module>
    guided_grads = GBP.generate_gradients(prep_img, target_class)
  File "guided_backprop.py", line 59, in generate_gradients
    model_output.backward(gradient=one_hot_output)
  File ".../torch/tensor.py", line 93, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph)
  File ".../torch/autograd/__init__.py", line 89, in backward
    ... allow_unreachable=True)  # allow_unreachable flag
TypeError: expected tuple, but hook returned 'Tensor'

From there you can trace it to torch/tensor.py and torch/autograd but in the end I couldn't find a good answer other than, its just what it is. Also, just to clarify, this hook function only affects backward pass and not forward pass, so ReLU(x) still returns modified_ReLU(x) but in the backward pass the gradients are affected.

from pytorch-cnn-visualizations.

Got it. Thanks. Your answer led me to read into Module class in Pytorch, where register_backward_hook actually explains it all.

One quick question. The vanilla_backprop and the guided_backprop seems to have much in common. You think refactoring could be of any use? I would be happy to do the work if you'd allow.

from pytorch-cnn-visualizations.

Thanks for the offer but many implementations are quite similar, and very simple but this is by design so that people can grasp these concepts and adapt to their own need without digging through hundreds of lines of code.

from pytorch-cnn-visualizations.