Gradient decent extracting data from simulations(returning numpy)

[ptrblck]

@Taboleiro

For the simulation program, I pass E, rho and eta as strings, ask for the simulation to run, and receive the output as a numpy array. Here is where the gradient path gets lost and does not fit or converge.

This is expected since Autograd won’t be able to track 3rd party operations. You would either need to stick to PyTorch tensors or you could implement custom autograd.Functions using e.g. numpy and implement the backward pass manually.