torch.nn.functional中softmax的作用及其参数说明

 参考：https://pytorch-cn.readthedocs.io/zh/latest/package_references/functional/#_1

class torch.nn.Softmax(input, dim)

或：

torch.nn.functional.softmax(input, dim)

 

对n维输入张量运用Softmax函数，将张量的每个元素缩放到（0,1）区间且和为1。Softmax函数定义如下：

参数：

　　dim:指明维度，dim=0表示按列计算；dim=1表示按行计算。默认dim的方法已经弃用了，最好声明dim，否则会警告：

UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument.

shape：

• 输入：(N, L)
• 输出：(N, L)

返回结果是一个与输入维度dim相同的张量，每个元素的取值范围在（0,1）区间。

例子：

import torch

from torch import nn
from torch import autograd

m = nn.Softmax()
input = autograd.Variable(torch.randn(2, 3))
print(input)
print(m(input))

返回：

(deeplearning) userdeMBP:pytorch user$ python test.py 
tensor([[ 0.2854,  0.1708,  0.4308],
        [-0.1983,  2.0705,  0.1549]])
test.py:9: UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument.
  print(m(input))
tensor([[0.3281, 0.2926, 0.3794],
        [0.0827, 0.7996, 0.1177]])

可见默认按行计算，即dim=1

 

更明显的例子：

import torch

import torch.nn.functional as F

x= torch.Tensor( [ [1,2,3,4],[1,2,3,4],[1,2,3,4]])

y1= F.softmax(x, dim = 0) #对每一列进行softmax
print(y1)

y2 = F.softmax(x,dim =1) #对每一行进行softmax
print(y2)

x1 = torch.Tensor([1,2,3,4])
print(x1)

y3 = F.softmax(x1,dim=0) #一维时使用dim=0，使用dim=1报错
print(y3)

返回：

(deeplearning) userdeMBP:pytorch user$ python test.py 
tensor([[0.3333, 0.3333, 0.3333, 0.3333],
        [0.3333, 0.3333, 0.3333, 0.3333],
        [0.3333, 0.3333, 0.3333, 0.3333]])
tensor([[0.0321, 0.0871, 0.2369, 0.6439],
        [0.0321, 0.0871, 0.2369, 0.6439],
        [0.0321, 0.0871, 0.2369, 0.6439]])
tensor([1., 2., 3., 4.])
tensor([0.0321, 0.0871, 0.2369, 0.6439])

因为列的值相同，所以按列计算时每一个所占的比重都是0.3333；行都是[1,2,3,4]，所以按行计算，比重结果都为[0.0321, 0.0871, 0.2369, 0.6439]

一维使用dim=1报错：

RuntimeError: Dimension out of range (expected to be in range of [-1, 0], but got 1)