Pytorch-区分nn.BCELoss()、nn.BCEWithLogitsLoss()和nn.CrossEntropyLoss() 的用法

详细理论部分可参考https://www.cnblogs.com/wanghui-garcia/p/10862733.html

BCELoss()和BCEWithLogitsLoss()的输出logits和目标labels的形状相同。

import torch 
import torch.nn as nn
import torch.nn.functional as F

m = nn.Sigmoid()

loss_f1 = nn.BCELoss()  
loss_f2 = nn.BCEWithLogitsLoss() 
loss_f3 = nn.CrossEntropyLoss() 

logits = torch.randn(3, 2)
labels = torch.FloatTensor([[0, 1], [1, 0], [1, 0]])

print(loss_f1(m(logits), labels))      #tensor(0.9314)，注意logits先被激活函数作用 
print(loss_f2(logits, labels))         #tensor(0.9314)

label2 = torch.LongTensor([1, 0, 0])
print(loss_f3(logits, label2))         #tensor(1.2842)

 如果label2也想变成labels，然后通过BCELoss进行计算的话，可以先转变成独热编码的形式：

encode = F.one_hot(label2, num_classes = 2)               #encode的值和labels一样，但是类型是LongTensor
print(loss_f1(m(logits), encode.type(torch.float32)))     #tensor(0.9314)