How to use torch.autograd.backward()

autograd
1

I have a network, for example input->conv1->bn1->relu->pool1. Now I have the gradient of pool1, and I want to calculate all the gradients of the network w.r.t the input, can it works: torch.autograd.backward(network, grad_tensors=pool1_gradient) ?

Hope someone can give answers, thanks!

2

Hi,

You can check the doc but backward takes the output of your forward (not the network) and you can give grad_tensors=pool1_gradient indeed.

3

Thanks for reply, I wonder if I use torch.autograd.backward(pool1_output, grad_tensors=pool1_gradient) , gradients of all layers w.r.t the input can be computed or just the gradient of pool1 is computed?

4

When you call .backward(), it will populate the .grad field for all the leaf Tensor (such as nn.Parameter) that were used to compute your output.