I have two net works A and B, the output is out = A(x1) + B(x2), it is a matrix. Now I want use out.backward(matrixj). I know there should be a maxtrxj, which has the same size with out, but I’m not sure if it is ok I just use a matrix with all element is 1.
It’s possible to do the backward with two networks. You have to create a loss function before you can do that though. Once you’ve calculated the loss you can do
See the pytorch examples for how it’s done with one network. It’s pretty much the same with two networks
Hi, thank you for your reply, but I think you maybe misunderstand me, I know we can use loss function for the back propagation, but what I’m doing now is use the matrix for the back propagation, I want use some thing like Jacobi matrix, I want to know if I can set all element is 1 in matrixj
Okay I see. Never done that myself. Good luck