參考:https://pytorch-cn.readthedocs.io/zh/latest/package_references/functional/#_1
class torch.nn.Softmax(input, dim)
或:
torch.nn.functional.softmax(input, dim)
對n維輸入張量運用Softmax函數,将張量的每個元素縮放到(0,1)區間且和為1。Softmax函數定義如下:
![](https://img.laitimes.com/img/9ZDMuAjOiMmIsIjOiQnIsIyZuBnLxUTN1UjNwQjMx0CNyIzNygzMwETOwQDM5EDMy0iMzAjN0QTMvwFNwkTMwIzLcJzMwYDN0EzLcd2bsJ2Lc12bj5ycn9Gbi52YugTMwIzZtl2Lc9CX6MHc0RHaiojIsJye.png)
參數:
dim:指明次元,dim=0表示按列計算;dim=1表示按行計算。預設dim的方法已經棄用了,最好聲明dim,否則會警告:
UserWarning: Implicit dimension choice for softmax has been deprecated. Change the call to include dim=X as an argument.
shape:
- 輸入:(N, L)
- 輸出:(N, L)
傳回結果是一個與輸入次元dim相同的張量,每個元素的取值範圍在(0,1)區間。
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))
傳回:
(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]])
可見預設按行計算,即dim=1
更明顯的例子:
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) #對每一列進行softmax
print(y1)
y2 = F.softmax(x,dim =1) #對每一行進行softmax
print(y2)
x1 = torch.Tensor([1,2,3,4])
print(x1)
y3 = F.softmax(x1,dim=0) #一維時使用dim=0,使用dim=1報錯
print(y3)
傳回:
(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])
因為列的值相同,是以按列計算時每一個所占的比重都是0.3333;行都是[1,2,3,4],是以按行計算,比重結果都為[0.0321, 0.0871, 0.2369, 0.6439]
一維使用dim=1報錯:
RuntimeError: Dimension out of range (expected to be in range of [-1, 0], but got 1)
轉載于:https://www.cnblogs.com/wanghui-garcia/p/10675588.html