Pytorch - Stack dimension must be exactly the same?

Question

In pytorch, given the tensors a of shape (1X11) and b of shape (1X11), torch.stack((a,b),0) would give me a tensor of shape (2X11)

However, when a is of shape (2X11) and b is of shape (1X11), torch.stack((a,b),0) will raise an error cf. "the two tensor size must exactly be the same".

Because the two tensor are the output of a model (gradient included), I can't convert them to numpy to use np.stack() or np.vstack().

Is there any possible solution for least GPU memory usage?

Answer 1

It seems you want to use torch.cat() (concatenate tensors along an existing dimension) and not torch.stack() (concatenate/stack tensors along a new dimension):

import torch

a = torch.randn(1, 42, 1, 1)
b = torch.randn(1, 42, 1, 1)

ab = torch.stack((a, b), 0)
print(ab.shape)
# torch.Size([2, 1, 42, 1, 1])

ab = torch.cat((a, b), 0)
print(ab.shape)
# torch.Size([2, 42, 1, 1])
aab = torch.cat((a, ab), 0)
print(aab.shape)
# torch.Size([3, 42, 1, 1])