Does my hook is correct?

autograd
1

I am using forward and backward hook in my pytorch densenet121 model.
I set requires_grad to False at the time of training.

for param in model.features.parameters():
    param.requires_grad = False

and I want to get gradient of last conv layer in Neural Network , for this i define hook like this:

model.features.denseblock4.denselayer16.conv2.register_backward_hook(h1)

but this is giving me empty list. But is i set requires_grad = True for the last layer then i get the gradient.

for param in model.features.denseblock4.denselayer16.conv2.parameters():
    param.requires_grad = True

I use gradient for gradCAM .

is this write way to doing or is there any other way to get gradient?

2

It’s expected that no gradients will be available in the hook, if you disable the gradient calculation.
register_backward_hook is also deprecated in favor of register_full_backward_hook, so you might want to change the usage.