pytorch 加权BCE

加权BCE_loss用来针对类别不均衡的二分类的语义分割，对于前景像素乘以一个比较大的权重，对背景像素乘以一个比较小的权重。
我知道的有两种实现方法：
假设网络预测为inputs，标签为target
第一种：
首先计算普通的BCEloss如下

import torch
import torch.nn as nn

inputs = torch.rand(1,1,3,3)
targets = torch.randint(1,3,(1,1,3,3))

loss = nn.BCELoss(reduction='none')(inputs,target.float())
print(loss)

Out[40]: 
tensor([[[[0.9652, 0.1823, 2.8128],
          [1.8717, 0.3409, 1.6916],
          [5.1682, 0.5676, 0.1342]]]])

然后加上权重，对于背景权重为0.3，前景权重为0.7

weight = torch.zeros_like(targets).float().cuda()
weight = torch.fill_(weight,0.3)
weight[target>0]=0.7
loss = torch.mean(weight*loss)
print(loss)

Out[68]: tensor(0.8863)

第二种方式：

inputs = torch.rand(1,1,3,3)
targets = torch.randint(1,3,(1,1,3,3))

weight = torch.zeros_like(targets).float().cuda()
weight = torch.fill_(weight,0.3)
weight[target>0]=0.7

loss = nn.BCELoss(weight=weight)(inputs,target.float())

Out[71]: tensor(0.8863)

加权交叉熵损失：
加权CE_loss和BCE_loss稍有不同
1.标签为long类型，BCE标签为float类型

2.当reduction为mean时计算每个像素点的损失的平均，BCE除以像素数得到平均值，CE除以像素对应的权重之和得到平均值。

inp = torch.rand(1,2,3,3)#因为是二分类，所以把inputs转换成两个通道,
inp[0,0]=1-inputs[0,0]
inp[0,1]=inputs[0,0]
loss = nn.CrossEntropyLoss(weight=torch.tensor([0.3,0.7]))(inp,target.squeeze(1))
print(loss)

Out[81]: tensor(0.8659)

参考https://blog.csdn.net/pain_gain0/article/details/106643607，写的很好