Pytorch常用函数、方法

• enumerate()

可获得索引和元素值，其中enumerate()第二个参数用来指定起始索引值，可省略，例

list = ["Python","Java","C"]

for index,key in enumerate(list):
    print(index,key)

• torch.randn()

输出正态分布随机数
example1：

import torch
import torch.nn as nn
input = torch.randn(1, 1, 3, 4)
print(input)

结果：

tensor([[[[ 0.5028, -0.7468,  1.8858,  0.1745],
          [ 0.8540,  0.0401,  1.4751,  0.9010],
          [-0.3230, -0.4141, -0.4215,  0.1705]]]])

example2：

import torch
import torch.nn as nn
input = torch.randn(1, 2, 3, 4)
print(input)

结果：

tensor([[[[ 1.5492,  0.9120,  0.9391, -0.2901],
          [-0.3356, -0.3431,  1.0347, -1.6674],
          [-0.1109, -0.1498, -1.2600,  0.1818]],

         [[ 1.8994,  0.0805, -2.7722, -1.1939],
          [ 1.4740, -0.2008, -0.1438, -1.1926],
          [-0.6315,  0.8516,  1.9624, -1.2148]]]])

• nn.Conv2d()

nn.Conv2d()前三个输入参数分别是(input_channels,output_chanels,kernel_sizes)
卷积，重点说明一下padding=N，就填充N个数(默认填充0)
example1：

import torch
import torch.nn as nn
input = torch.randn(1, 1, 3, 4)
print(input)
m = nn.Conv2d(1,1,1,stride=1,bias=False)
print(m(input))
m1 = nn.Conv2d(1,1,1,stride=1,bias=False,padding=1)
print(m1(input))
m2 = nn.Conv2d(1,1,1,stride=1,bias=False,padding=2)
print(m2(input))

结果：

tensor([[[[ 1.2922, -1.6056, -0.2292, -1.1778],   
          [-1.1310, -1.9764, -1.2235, -0.5288],   
          [ 1.5305, -0.1229, -1.3054,  1.3235]]]])
tensor([[[[ 0.3306, -0.4108, -0.0586, -0.3014],
          [-0.2894, -0.5057, -0.3131, -0.1353],
          [ 0.3916, -0.0314, -0.3340,  0.3386]]]],
       grad_fn=<ThnnConv2DBackward>)
tensor([[[[ 0.0000,  0.0000,  0.0000,  0.0000,  0.0000,  0.0000],   
          [ 0.0000,  0.7538, -0.9366, -0.1337, -0.6871,  0.0000],
          [ 0.0000, -0.6598, -1.1529, -0.7137, -0.3085,  0.0000],
          [ 0.0000,  0.8928, -0.0717, -0.7615,  0.7720,  0.0000],
          [ 0.0000,  0.0000,  0.0000,  0.0000,  0.0000,  0.0000]]]],
       grad_fn=<ThnnConv2DBackward>)
tensor([[[[ 0.0000,  0.0000,  0.0000,  0.0000,  0.0000,  0.0000,  0.0000,
            0.0000],
          [ 0.0000,  0.0000,  0.0000,  0.0000,  0.0000,  0.0000,  0.0000,
            0.0000],
          [ 0.0000,  0.0000, -0.3415,  0.4243,  0.0606,  0.3113,  0.0000,
            0.0000],
          [ 0.0000,  0.0000,  0.2989,  0.5223,  0.3233,  0.1397,  0.0000,
            0.0000],
          [ 0.0000,  0.0000, -0.4045,  0.0325,  0.3450, -0.3498,  0.0000,
            0.0000],
          [ 0.0000,  0.0000,  0.0000,  0.0000,  0.0000,  0.0000,  0.0000,
            0.0000],
          [ 0.0000,  0.0000,  0.0000,  0.0000,  0.0000,  0.0000,  0.0000,
            0.0000]]]], grad_fn=<ThnnConv2DBackward>)

• permute()

调换Tensor中各维度的顺序
example1：

import torch

x = torch.empty([6, 7, 8, 9])
print(x.size())
x = x.permute([0, 1, 3, 2])
print(x.size())

输出：

torch.Size([6, 7, 8, 9])
torch.Size([6, 7, 9, 8])

调换前tensor第三维度有9个元素，调换后第三维度变成8个元素(和第四维度的发生调换)，第四维度有9个元素

• view()

view的作用是reshape(),要求tensor是连续的，一般用法是.contiguous().view(n, -1,…)
example1：

扫描二维码关注公众号，回复： 12828393 查看本文章

a = torch.arange(1, 17)  # a's shape is (16,)
 
print(a.view(4, 4)) # output below
print(a.view(2, 2, 4)) # output below

结果：

tensor([[ 1,  2,  3,  4],
        [ 5,  6,  7,  8],
        [ 9, 10, 11, 12],
        [13, 14, 15, 16]])
[torch.FloatTensor of size 4x4]

tensor([[[ 1,  2,  3,  4],
         [ 5,  6,  7,  8]],
 
        [[ 9, 10, 11, 12],
         [13, 14, 15, 16]]])
[torch.FloatTensor of size 2x2x4]

a.view(2, 2, 4)表示reshape成2个2*4矩阵，若写成a.view(n,-1,4)则表示reshape成n个x行4列的矩阵(-1会自动匹配有多少行)