I’ve seen that one can implement their own gradients for a given function inheriting from torch.autograd.Function
and implementing the forward
and backward
static methods.
I would like to know how to have access to the equivalent to the def backward(ctx, grad_output)
method for an arbitrary function implemented in pytorch.
According to the Pytorch documentation:
Every operation performed on
Tensor
s creates a new function object, that performs the computation, and records that it happened.
So I believe that this should be possible to compute automatically.