Is layer_activation (register_forward_hook) the same as gradient?

1

I was wondering if the intermediate layer output initialised by register_forward_hook is the same as the gradient of the intermediate output wrt the image?

2

No, the forward activations are not the gradients unless I’m misunderstanding your question. This tutorial could be useful as it shows which values are used during the forward and backward pass.

3

Thanks!

My end goal is to calculate the gradient of layer activations/intermediate layer outputs wrt the input image. However, because of a next step I know that the end result needs to be the same size of my layer activations/intermediate layer outputs array. With the following code I seem to be getting an end result that is the same size as my input image. Hence the confusion with gradients and layer activations. This is the code that I am using:


 model.eval()
    tcav = {}
    for ind, (img, label) in enumerate(loader):
        img.requires_grad=True
        img = img.to(device, dtype=torch.float)
        output = model(img)
        
        layer_activation = activation[L] #activation array of layer L
        loss = -torch.mean(layer_activation)
        loss.backward(retain_graph=True)
        
        gradients = tuple_of_tensors_to_tensor(torch.autograd.grad(loss, img))[0]
        grads = normalize(gradients.cpu().detach().numpy().ravel())
        tcav[ind] = layer_activation#grads

Do you perhaps know what I need to do?