Multiple functions sharing some gradients about autograd.jl HOT 1 CLOSED

I just realized that in a simple case like this one can just take the gradient of discriminator output separately, and use the chain rule to avoid extra computation. It won't be pretty when the functions are more entangled, but I suppose it's the best option.

I just want to add that the chain rule method doesn't work when the shared computation does not result in a scalar. I am wondering if there is a more general way of doing these calculations.

from autograd.jl.