Pytorch中的维度变换，reshape、permute、view

程序中常常会涉及到张量的维度变换，为区别常用变换函数reshape、permute、view的不同，使用以下代码直观说明。
先对二维张量进行变换：

import torch
c=torch.randn(2,3)
print("c:")
print(c)
c1=torch.reshape(c,(3,2))
print("reshape:")
print(c1)
c2=c.permute(1,0)
print("permute:")
print(c2)
c3=c.view(3,2)
print("view:")
print(c3)
c4=c.transpose(1,0)
print("transpose:")
print(c4)

代码结果：

c:
tensor([[-0.1440,  0.0783,  0.1932],
        [-0.3421, -0.3892,  1.5916]])
reshape:
tensor([[-0.1440,  0.0783],
        [ 0.1932, -0.3421],
        [-0.3892,  1.5916]])
permute:
tensor([[-0.1440, -0.3421],
        [ 0.0783, -0.3892],
        [ 0.1932,  1.5916]])
view:
tensor([[-0.1440,  0.0783],
        [ 0.1932, -0.3421],
        [-0.3892,  1.5916]])
transpose:
tensor([[-0.1440, -0.3421],
        [ 0.0783, -0.3892],
        [ 0.1932,  1.5916]])

先对三维张量进行变换（）

import torch
c=torch.randn(2,3,4)
print("c:")
print(c)
c1=torch.reshape(c,(4,3,2))
print("reshape:")
print(c1)
c2=c.permute(2,1,0)
print("permute:")
print(c2)
c3=c.view(4,3,2)
print("view:")
print(c3)

运行结果：

c:
tensor([[[-1.0615,  0.4882,  0.8409,  0.2768],
         [ 0.9477, -0.3273, -1.0569, -0.8247],
         [ 1.6808,  0.8768, -0.0501,  0.6021]],

        [[-0.7654, -0.0558,  0.6285, -1.5840],
         [ 0.0180,  0.0347,  0.8596, -1.1917],
         [ 3.1598, -0.2194,  0.9064,  0.2113]]])
reshape:
tensor([[[-1.0615,  0.4882],
         [ 0.8409,  0.2768],
         [ 0.9477, -0.3273]],

        [[-1.0569, -0.8247],
         [ 1.6808,  0.8768],
         [-0.0501,  0.6021]],

        [[-0.7654, -0.0558],
         [ 0.6285, -1.5840],
         [ 0.0180,  0.0347]],

        [[ 0.8596, -1.1917],
         [ 3.1598, -0.2194],
         [ 0.9064,  0.2113]]])
permute:
tensor([[[-1.0615, -0.7654],
         [ 0.9477,  0.0180],
         [ 1.6808,  3.1598]],

        [[ 0.4882, -0.0558],
         [-0.3273,  0.0347],
         [ 0.8768, -0.2194]],

        [[ 0.8409,  0.6285],
         [-1.0569,  0.8596],
         [-0.0501,  0.9064]],

        [[ 0.2768, -1.5840],
         [-0.8247, -1.1917],
         [ 0.6021,  0.2113]]])
view:
tensor([[[-1.0615,  0.4882],
         [ 0.8409,  0.2768],
         [ 0.9477, -0.3273]],

        [[-1.0569, -0.8247],
         [ 1.6808,  0.8768],
         [-0.0501,  0.6021]],

        [[-0.7654, -0.0558],
         [ 0.6285, -1.5840],
         [ 0.0180,  0.0347]],

        [[ 0.8596, -1.1917],
         [ 3.1598, -0.2194],
         [ 0.9064,  0.2113]]])

可以看出，reshape只是对形状做了一个变化，而permute，transpose则是对轴进行了一个互换。