pytorch nn.LogSoftmax

应用

# softmax层
# m = nn.Softmax(dim=1)
# input = torch.randn(2, 3)
# output = m(input)

m = nn.LogSoftmax()
input = torch.randn(2, 3)
output = m(input)

import torch
input=torch.randn(2,2)
input # 每一行代表一个样本，每一列代表该样本属于某一类
---------------------------
tensor([[-0.3549,  1.1395],
        [ 2.1887,  1.0817]])
---------------------------
soft = torch.nn.Softmax(dim=1) # 横向计算softmax
logsoft = torch.nn.LogSoftmax(dim=1) # 横向计算softmax
soft(input) # 将输出转化为概率
---------------------------
tensor([[0.1833, 0.8167],
        [0.7516, 0.2484]])
---------------------------
logsoft(input) # 等价于torch.log(soft_input(input))
-----------------------------
tensor([[-1.6968, -0.2024],
        [-0.2856, -1.3926]])
-----------------------------

API

$LogSoftmax(x_i)=log(\frac{exp(x_i)}{{\sum }_{j}exp(x_j)})$
在softmax层外加了log函数

CLASS torch.nn.LogSoftmax(dim: Optional[int] = None)

参考

https://pytorch.org/docs/stable/generated/torch.nn.LogSoftmax.html?highlight=nn%20logsoftmax#torch.nn.LogSoftmax