Inference of NN gradient w.r.t. inputs

[julsanmer]

Hello,

I'd like to inquiry if there is a way or workaround to do inference in ONNXRuntime with the neural network gradient w.r.t. inputs. I get an error when exporting torch.autograd.grad operator. In the field of physics-informed neural networks, the desired output is oftentimes the NN gradient.

For now, I'm using finite differentiation of the NN but it is clearly a sub-optimal solution.