torch.Tensor.transpose()

first dimension index

second dimension index

torch.Tensor.transpose() is not just sugar syntax, but rather

torch tensor.permute() is the sugar syntax. While transpose explicitly specifies two dimensions to swap,

Pytorch tensor.T completely reverses the shape Size array, essentially a version of permute with reversed size. tensor.mT is specifically for 2D and only allows 2D tensors.


import torch

# .T reverses all dimensions
x = torch.randn(2, 3, 4)
print(x.shape)  # torch.Size([2, 3, 4])
print(x.T.shape)  # torch.Size([4, 3, 2])

# .mT is for 2D only
y = torch.randn(3, 5)
print(y.mT.shape)  # torch.Size([5, 3])

# transpose swaps two specific dimensions
print(x.transpose(0, 2).shape)  # torch.Size([4, 3, 2])

# permute reorders dimensions in any order
print(x.permute(2, 0, 1).shape)  # torch.Size([4, 2, 3])
print(x.permute(2, 1, 0).shape)  # torch.Size([4, 3, 2]) - same as .T

torch.Tensor.transpose()

Backlinks

Recommendations