Torchconnector fails in backward prop

[qiskit-community]

Torchconnector fails in backward prop

[adekusar-drl]

@woodsp-ibm Interesting, this may happen if input_grad.shape has a zero element. A solution off the top of my head is just to add an additional if input_grad is not None before the einsum, but I'm not sure. Is there a test case to reproduce the issue?

[woodsp-ibm]

Indeed doing the einsum conditionally like that, which is how the failure was avoided so further progress could be made, but since it changed input_grad to None, which is then what is returned, I was not so sure. I don't have a small test case at present but I can see if I can get something even the failing code itself.

[IceKhan13]

I think, those cases should be handled in NeuralNetwork validate methods (validate inputs, weights and backwards).

https://github.com/Qiskit/qiskit-machine-learning/blob/00831061ca0206af42e131fcb9b605c4acd404fa/qiskit_machine_learning/neural_networks/neural_network.py#L217-L223

Looks like something is zeroing one of the dims in implementation of _backward.

I can look at it

[adekusar-drl]

@IceKhan13 Sure, you can take a look.