I am trying to compute hessian matrix of input. It is straightforward to get the 1st derivative of input. But is there a way to compute 2nd order gradients of loss w.r.t to input in pytorch?
You can compute Hessian-Vector-Products, the trick is to pass build_graph=True
to torch.autograd.grad
or x.backward
. Then you can use gradients in computation and y.backward
or torch.autograd.grad
again will do the right thing.
Best regards
Thomas