python pytorch中 .view()函数讲解

在使用pytorch定义神经网络时，经常会看到类似如下的.view()用法，这里对其用法做出讲解与演示。

一、普通用法 (手动调整size)

view()相当于reshape、resize，重新调整Tensor的形状。

import torch
a1 = torch.arange(0,16)
print(a1)

tensor([ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15])

a2 = a1.view(8, 2)
a3 = a1.view(2, 8)
a4 = a1.view(4, 4)

print(a2)
print(a3)
print(a4)

tensor([[ 0,  1],
        [ 2,  3],
        [ 4,  5],
        [ 6,  7],
        [ 8,  9],
        [10, 11],
        [12, 13],
        [14, 15]])
tensor([[ 0,  1,  2,  3,  4,  5,  6,  7],
        [ 8,  9, 10, 11, 12, 13, 14, 15]])
tensor([[ 0,  1,  2,  3],
        [ 4,  5,  6,  7],
        [ 8,  9, 10, 11],
        [12, 13, 14, 15]])

二、特殊用法：参数-1 (自动调整size)

view中一个参数定为-1，代表自动调整这个维度上的元素个数，以保证元素的总数不变。

import torch
a1 = torch.arange(0,16)
print(a1)

tensor([ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15])

a2 = a1.view(-1, 16)
a3 = a1.view(-1, 8)
a4 = a1.view(-1, 4)
a5 = a1.view(-1, 2)
a6 = a1.view(4*4, -1)
a7 = a1.view(1*4, -1)
a8 = a1.view(2*4, -1)

print(a2)
print(a3)
print(a4)
print(a5)
print(a6)
print(a7)
print(a8)

tensor([[ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15]])
tensor([[ 0,  1,  2,  3,  4,  5,  6,  7],
        [ 8,  9, 10, 11, 12, 13, 14, 15]])
tensor([[ 0,  1,  2,  3],
        [ 4,  5,  6,  7],
        [ 8,  9, 10, 11],
        [12, 13, 14, 15]])
tensor([[ 0,  1],
        [ 2,  3],
        [ 4,  5],
        [ 6,  7],
        [ 8,  9],
        [10, 11],
        [12, 13],
        [14, 15]])
tensor([[ 0],
        [ 1],
        [ 2],
        [ 3],
        [ 4],
        [ 5],
        [ 6],
        [ 7],
        [ 8],
        [ 9],
        [10],
        [11],
        [12],
        [13],
        [14],
        [15]])
tensor([[ 0,  1,  2,  3],
        [ 4,  5,  6,  7],
        [ 8,  9, 10, 11],
        [12, 13, 14, 15]])
tensor([[ 0,  1],
        [ 2,  3],
        [ 4,  5],
        [ 6,  7],
        [ 8,  9],
        [10, 11],
        [12, 13],
        [14, 15]])