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?
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?