Is there any special reason to work with forward_hook instead of backward_hook to get the gradients?

[tldrafael]

Hi, thanks for the amazing work!

I am trying to implement a plain version of the GradCAM by myself. However, I tried to use register_module_full_backward_hook instead of register_forward_hook, I got different values for many elements of the gradient tensor, although around 70% of tensor is equal.

pytorch-grad-cam/pytorch_grad_cam/activations_and_gradients.py

Line 17 in 60cf39d

target_layer.register_forward_hook(self.save_gradient))

I wonder if there is anything that I did not figure out. Is it expected they show different results?