Unexpected output while printing gradients of a layer when backward hook is applied on it

Hi everyone!
I’m trying to understand the working of backward hook and thus playing around with it to observe and learn.
I applied a backward hook on a simple covnet. The hook looks something like this:

def hook_function(m,i,o):
    estimated_gradient = torch.ones_like(i[0])
    estimated_gradient = torch.unsqueeze(estimated_gradient, dim=0)
    return estimated_gradient

Here I’m return a tensor of 1s as the input gradient for the next layer to use. I print these values after doing the forward propagation (but before doing loss.backwards and optimizer.step)
The code looks something like this -

#print(list(model.parameters())[6].shape) # I used this to identify the particular layer I want to print

This prints the grads of the last layer in my covnet . The code where I apply these hooks on the layers of my model is as follows (I am applying these hooks on all the linear layers of my covnet-

def register_backward_hook_for_(Model):
    target_modules = []
    for m in Model.modules():
        if isinstance(m,nn.Linear):
    # print(target_modules)
    for modules in target_modules:

Expected output - I expected these values to be 1, as I did not perform any backward computation (printing it before my loss.backward call)
Actual output - Real valued outputs

What is going wrong? Or am I missing something about the working of these hooks?

I don’t quite understand this statement as backward hooks will be called during the .backward() call.
Could you explain what exactly you are printing? If no backward() call was used, all .grad attributes would be set to None.

Thank you for the reply sir !
Now I get why I’m not able to print 1s as the output of gradients for any particular layer. I did not realize this fact and also that the values of gradients being printed would be the ones stored in these tensors from the previous epoch.
I really appreciate you helping me out, thank you!

Ah OK. Thanks for the update and good to hear the issue seems to be solved! :slight_smile:

1 Like