The documentation says that autograd will take care of backpropagation when defining the loss.
However, are there any conditions that we need to consider when defining a custom loss? For example, should the loss always be differentiable?
1 Like