Pytorch-损失函数-NLLLoss

https://blog.csdn.net/weixin_40476348/article/details/94562240

常用于多分类任务，NLLLoss 函数输入 input 之前，需要对 input 进行 log_softmax 处理，即将 input 转换成概率分布的形式，并且取对数，底数为 e

class torch.nn.NLLLoss(weight=None, size_average=None, ignore_index=-100, 
					   reduce=None, reduction='mean')

计算公式：loss(input, class) = -input[class]
公式理解：input = [-0.1187, 0.2110, 0.7463]，target = [1]，那么 loss = -0.2110
个人理解：感觉像是把 target 转换成 one-hot 编码，然后与 input 点乘得到的结果
代码理解：

import torch
import torch.nn as nn
import torch.nn.functional as F
torch.manual_seed(2019)
output = torch.randn(1, 3)  # 网络输出

target = torch.ones(1, dtype=torch.long).random_(3)  # 真实标签

print(output)

print(target)
# 直接调用

loss = F.nll_loss(output, target)

print(loss)
# 实例化类

criterion = nn.NLLLoss()

loss = criterion(output, target)

print(loss)
"""

tensor([[-0.1187,  0.2110,  0.7463]])

tensor([1])

tensor(-0.2110)

tensor(-0.2110)

"""

如果 input 维度为 M x N，那么 loss 默认取 M 个 loss 的平均值，reduction='none' 表示显示全部 loss

import torch
import torch.nn as nn
import torch.nn.functional as F
torch.manual_seed(2019)
output = torch.randn(2, 3)  # 网路输出

target = torch.ones(2, dtype=torch.long).random_(3)  # 真实标签

print(output)

print(target)
# 直接调用

loss = F.nll_loss(output, target)

print(loss)
# 实例化类

criterion = nn.NLLLoss(reduction='none')

loss = criterion(output, target)

print(loss)
"""

tensor([[-0.1187,  0.2110,  0.7463],

[-0.6136, -0.1186,  1.5565]])

tensor([2, 0])

tensor(-0.0664)

tensor([-0.7463,  0.6136])

"""