AutoDiff for matrix exponential

autograd
1

I got the following RuntimeError for the matrix exponential: element 0 of tensors does not require grad and does not have a grad_fn. I don’t know how to make it work. Does anyone know where the problem is?

def test(x):
exp_x = torch.tensor(expm(x.detach().numpy()))
return torch.sum(exp_x)

x = torch.eye(3,3)
x = x.requires_grad_()
y = test(x)
y.backward()

2

Since you are detaching the computation graph from your tensor manually, x won’t get gradients anymore.
I assume you are detaching x to call expm from scipy.
You could try to implement the Pade approximation in PyTorch or write an autograd.Function with a custom backward method.

3

Thanks for your reply. I have used a different algorithm to do my work instead of using the matrix exponential.