Say there is function F . which cannot be computed exactly but it’s gradient can be perfectly defined can I define function s.t I tell the autograd , how gradient should be computed instead of autograd figuring about by computation graph. So my requirement is as following,
function F(x) : #cannot be computed exactly approx = some_decent_approx_of_F(x) return approx
function Grad_of_F(x): #can be computed exactly grad = exact_grad_of_F(x) return grad
so instead of gradient computed based on computational graph of F. I want to explicitly say use Grad_of_F for computation of gradient in backward pass ? It would really helpful if someone could point to how to do it. Thanks In advance ?