Pytorch Tensor形状变换

之前一直知道Pytorch、numpy之类有好几种变换维度的方法，但是之前了解的不深，今天总结一下。

Reshape

Pytorch在reshape的时候，是按照行存储的方式进行的。

import torch

In [16]: x = torch.randn(2, 3, 4)

In [17]: x
Out[17]:
tensor([[[ 0.2921, -0.1806, -1.0838, -0.6770],
         [-0.7797,  0.2614, -0.5380,  1.8941],
         [-0.7261,  0.8209,  0.0286,  0.0997]],

        [[ 0.5549,  0.9036,  0.8790, -0.5776],
         [ 0.3745,  0.6963, -0.3445, -0.0022],
         [-0.5991,  1.4639, -0.5396, -0.1702]]])

In [19]: torch.reshape(x, (-1, 2))
Out[19]:
tensor([[ 0.2921, -0.1806],
        [-1.0838, -0.6770],
        [-0.7797,  0.2614],
        [-0.5380,  1.8941],
        [-0.7261,  0.8209],
        [ 0.0286,  0.0997],
        [ 0.5549,  0.9036],
        [ 0.8790, -0.5776],
        [ 0.3745,  0.6963],
        [-0.3445, -0.0022],
        [-0.5991,  1.4639],
        [-0.5396, -0.1702]])

permute

permute会将tensor维度进行调整。

In [24]: y = x.permute(1, 2, 0)

In [25]: y
Out[25]:
tensor([[[ 0.2921,  0.5549],
         [-0.1806,  0.9036],
         [-1.0838,  0.8790],
         [-0.6770, -0.5776]],

        [[-0.7797,  0.3745],
         [ 0.2614,  0.6963],
         [-0.5380, -0.3445],
         [ 1.8941, -0.0022]],

        [[-0.7261, -0.5991],
         [ 0.8209,  1.4639],
         [ 0.0286, -0.5396],
         [ 0.0997, -0.1702]]])

In [26]: y.size()
Out[26]: torch.Size([3, 4, 2])

In [29]: x[:, 1, 1]
Out[29]: tensor([0.2614, 0.6963])

In [31]: y[1, 1, :]
Out[31]: tensor([0.2614, 0.6963])