BCEWithLogitsLoss(predict,target) 和BCELoss(predict,target) 的区别:

BCEWithLogitsLoss会对predict进行sigmoid处理；BCELoss不会对predict进行sigmoid处理；

#%%

import torch 
import torch.nn as nn

#%% md

BCEWithLogitsLoss损失函数

#%% 产生predict,产生target
N = 2
C = 1
H = 2
W = 2
predict = torch.arange(8,dtype=torch.float32).view([N,C,H,W])
predict = torch.sigmoid(predict) 
target = torch.arange(start=1,end=9,step=1,dtype=torch.float32).view([N,C,H,W])
target  = torch.sigmoid(target)  
print('predict:',predict)
print('target:',target)

#%% 利用pytorch计算BCEWithLogitsLoss
loss = torch.nn.BCEWithLogitsLoss()
print('Pytorch计算结果:BCEWithLogitsLoss:',loss(predict,target))

#%% 对predict做sigmoid+BCELoss
predict = torch.sigmoid(predict) 
loss = torch.nn.BCELoss()
print('Pytorch计算结果:sigmoid+BCELoss:',loss(predict,target))

结果：

Pytorch计算结果:BCEWithLogitsLoss: tensor(0.3864)

Pytorch计算结果:sigmoid+BCELoss: tensor(0.3864)