torch. Usage and comparison of unsqueeze() squeeze() expand() repeat()

1. torch.unsqueeze()

torch.unsqueeze(input, dim, out=None)

• effect ： Expand dimensions , Returns a new tensor , Insert a dimension into the given location of the input 1
• Be careful ： Return tensor and input tensor share memory , So changing the content of one will change the other .
  If dim Negative , It will be transformed into dim+input.dim()+1
• Parameters :
  tensor (Tensor) – Input tensor
  dim (int) – Insert the index of the dimension
  out (Tensor, optional) – The result tensor

2. torch.squeeze()

torch.squeeze(input, dim=None, out=None)

• effect ： Dimension reduction , Will enter... In the tensor shape 1 To remove and return to . If the input is in the form of (A×1×B×1×C×1×D), So the output shape is ： (A×B×C×D)
  When given dim when , So squash is only on a given dimension . for example , The input shape is : (A×1×B), squeeze(input, 0) It's going to keep the tensor constant , Only with squeeze(input, 1), The shape will change (A×B).
• Be careful ： Return tensor and input tensor share memory , So changing the content of one will change the other .
• Parameters :
  input (Tensor) – Input tensor
  dim (int, optional) – If a given , be input It only squeezes in a given dimension
  out (Tensor, optional) – Output tensor

3. torch.expand()

• effect ： expand() Function can broadcast tensors to new shapes .
• Be careful ： Only dimension values of 1 Expand the dimension of , And extended Tensor No new memory will be allocated , Just create a new view based on the original and return to ; Dimension values that do not need to be extended remain unchanged , Dimension for -1 That is, unchanged .

4. torch.repeat()

• effect ： and expand() Works in a similar way , All will tensor Broadcast to new shapes .
• Be careful ： Dimensions are not allowed -1,1 Is unchanged .

The following is an example of specific function usage .

a = torch.rand((2,1,3,1))
# torch.Size([2,1,3,1])
b = torch.unsqueeze(a, 1)
# torch.Size([2,1,1,3,1])
c = torch.unsqueeze(a, 0)
# torch.Size([1,2,1,3,1])
d = torch.unsqueeze(a, -1)
# torch.Size([2,1,3,1,1])
e = a.torch.expand((2,3,3,2))
# torch.Size([2,3,3,2])
f = a.torch.expand((2,-1,3,2))
# torch.Size([2,1,3,2])
g = a.torch.repeat((1,1,1,2))
# torch.Size([2,1,3,2])
h = torch.squeeze(a)
# torch.Size([2,3])
i = torch.squeeze(a, 0)
# torch.Size([2,1,3,1])
j = torch.squeeze(a, 1)
# torch.Size([2,3,1])