Why don't pooling functions implement `backward()`?

1

I’ve been trying to implement an extension to autograd for index-based pooling (rather than spatial pooling). However, when I implement backward() and then try to call it in my tests, I get the error:

RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

Now that I check the PyTorch built-ins too, I see the same error pop up.

import torch
x = torch.rand(2,10,5)
pool = torch.nn.MaxPool1d(5)
out = pool(x)

out.backward()
>>>RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

Why is there no gradient? I thought that the differentiability in max-pooling is achieved by backpropagating the gradient to each max element from the input, and none elsewhere. Can someone please explain how this works in PyTorch?

2

Hi,

The thing is that there is nothing that requires gradients in your graph. So there is nothing to backward to. You should add require_grad=True to the input.