How to use and preserve the grad history unchanged?

(Deeply) #1

Simply, I have a few variables, say x1, x2, x3, with requires_grad set to True. Using torch.Tensor resulted in removing all the grad history,

ipdb> x1
tensor(0.6684, device='cuda:0', grad_fn=<L1LossBackward>)

ipdb> x2
tensor(0.7662, device='cuda:0', grad_fn=<L1LossBackward>)

ipdb> torch.Tensor((x1, x2))
tensor([0.6684, 0.7662])

I need to concatenate the variables into one tensor and preserve the grad history, so that I can use backward pass later.

I have tried using torch.Tensor( (x1, x2, x3) ) but all the grad history has been lost.
Using,x2,x3)) did not work too, throwing the error message “RuntimeError: zero-dimensional tensor (at position 0) cannot be concatenated” .

One way around this, is to define the tensor
z = torch.ones(3, dtype=torch.float32, requires_grad=True).to(device)

then, multiply z by x

z[0] = z[0]*x1
z[1] = z[1]*x2
z[2] = z[2]*x3

Any, ideas, alternatives?


Could you try to use torch.stack instead?

1 Like