[resolved] How does one create copy of tensors in PyTorch?

1

I was trying to create new copies (so that python was pointing at different objects and not just updating the same copy across multiple namespaces). I tried it by getting the actual numpy data from a tensor and using that to create a brand new tensor but that didn’t work. Why did it not work? How does one do this?

>>> y

 3  3
 3  3
 3  3
[torch.FloatTensor of size 3x2]

>>> yv.data = torch.FloatTensor( y.numpy() )
>>> yv
Variable containing:
 3  3
 3  3
 3  3
[torch.FloatTensor of size 3x2]

>>> y

 3  3
 3  3
 3  3
[torch.FloatTensor of size 3x2]

>>> yv.data.fill_(5)

 5  5
 5  5
 5  5
[torch.FloatTensor of size 3x2]

>>> yv
Variable containing:
 5  5
 5  5
 5  5
[torch.FloatTensor of size 3x2]

>>> y

 5  5
 5  5
 5  5
[torch.FloatTensor of size 3x2]
4 Likes
2

clone maybe?

need to check it out…

this seems to be Variables but I wanted just Tensor only…

2 Likes
3

ok that works:

>>> yv.data = y.clone()
>>> y

 5  5
 5  5
 5  5
[torch.FloatTensor of size 3x2]

>>> yv
Variable containing:
 5  5
 5  5
 5  5
[torch.FloatTensor of size 3x2]

>>> yv.data.fill_(10)

 10  10
 10  10
 10  10
[torch.FloatTensor of size 3x2]

>>> y

 5  5
 5  5
 5  5
[torch.FloatTensor of size 3x2]

>>> yv
Variable containing:
 10  10
 10  10
 10  10
[torch.FloatTensor of size 3x2]
10 Likes
4

Note that indexing with a (non-zero dimensional) tensor also results in a copy.

(for zero-dim tensors see this merged commit)

5

You should also detach the tensor from the computational graph if it requires grad else gradients will be calculated for both that may lead to OOM issues.

x = torch.tensor([1.,2.,3.], requires_grad=True)
y = x.detach().clone()
y.requires_grad = True