Hi. For a related project, I need to access the gradients with respect to weights. In particular, the weights which are computed as: (w = w - learning_rate * grad
) can be accessed using layer.weights
. However, at every layer, i need to access the grad
that is used to update the weights, and not the weights. Is there any way to do this?
Hi,
After calling .backward()
, your weights will have a .grad
field that contain a Tensor with the gradients for this Tensor. So layer.weights.grad
in your case.