在CE和MSE损失函数中使用置信度的方法

            以下是在一个半监督情景中  weak_output_ul为弱扰动出来的logits   ,strong_output_ul为强扰动出来的logits  两者尺寸都可看作[8,2,256,256]

CE:

            weak_x_ul = self.encoder(A_ul, B_ul)
            weak_output_ul = self.main_decoder(weak_x_ul)
            weak_targets = F.softmax(weak_output_ul.detach(), dim=1)

            strong_x_ul = self.encoder(s_A_ul, s_B_ul)
            strong_output_ul = self.main_decoder(strong_x_ul)

            #我们先得到伪标签 [8,256,256]
            pseudo_labels = torch.argmax(weak_targets, dim=1)        

            #得到概率最大值和位置 [8,256,256]    

            max_probs, _ = torch.max(weak_targets, dim=1)

               #得到Ture False矩阵  [8,256,256]

            confidence_mask = max_probs > 0.95

             #.long()是输入要求    reduction=none是为了不平均  得到概率举证
            loss = F.cross_entropy(strong_output_ul, pseudo_labels.long(), reduction='none')
            
            loss = loss * confidence_mask.float()

            #计算需要的平均损失

            loss_unsup = loss.sum() / confidence_mask.sum()

MSE:

import torch
import torch.nn.functional as F


weak_targets=torch.rand(8,2,256,256)
weak_targets=torch.softmax(weak_targets,dim=1)

prob,_=torch.max(weak_targets,dim=1)
confidence_mask=prob>0.95

strong_output_ul=torch.rand(8,2,256,256)

strong_output_ul = F.softmax(strong_output_ul, dim=1) 


mse_loss = F.mse_loss(strong_output_ul, weak_targets, reduction='none')

# 应用置信度掩码
print(mse_loss.size())
mse_loss = mse_loss * confidence_mask.unsqueeze(1).float()  # 确保confidence_mask在应用前与mse_loss的形状匹配

loss_unsup = mse_loss.sum(dim=[1, 2, 3]) / confidence_mask.sum()

# 计算最终的平均损失
loss_unsup = loss_unsup.mean()