pytorch的diceloss

### PyTorch中的Dice Loss实现 在PyTorch中，可以自定义Dice Loss来处理分割任务。下面是一个基于Sigmoid激活函数的二分类Dice Loss实现[^1]： ```python import torch import torch.nn as nn import torch.nn.functional as F class DiceLoss(nn.Module): def __init__(self, smooth=1e-6): super(DiceLoss, self).__init__() self.smooth = smooth def forward(self, inputs, targets): # Flatten label and prediction tensors inputs = inputs.view(-1) targets = targets.view(-1) intersection = (inputs * targets).sum() dice_coefficient = (2.*intersection + self.smooth) / \ (inputs.sum() + targets.sum() + self.smooth) return 1 - dice_coefficient ``` 对于多类别分割问题，则通常采用Softmax作为最后一层激活函数并调整损失计算方式以适应多个类别的预测输出: ```python class MultiClassDiceLoss(nn.Module): def __init__(self, weight=None, size_average=True): super(MultiClassDiceLoss, self).__init__() def forward(self, inputs, target,smooth=1): N = target.size(0) C = inputs.size(1) # Predicted probabilities for each class probs = F.softmax(inputs,dim=1) mtarget = one_hot(target,C) # Reshape to have a separate dimension per channel/class mtarget = mtarget.permute(0,3,1,2).contiguous().view(N,C,-1) probs = probs.view(N,C,-1) intersection = (probs*mtarget).sum(dim=-1) cardinality = probs.sum(dim=-1)+mtarget.sum(dim=-1) dice_loss_per_class = ((2. * intersection + smooth)/(cardinality + smooth)).mean() return 1-dice_loss_per_class.mean() def one_hot(labels,num_classes): """ Convert integer labels into one-hot vectors. Args: labels(tensor): input tensor of shape [N,H,W]. num_classes(int): number of classes. Returns: Tensor with the same batch size but channels equaling `num_classes`. """ n_labels = labels.long().unsqueeze_(dim=1) one_hot = torch.FloatTensor(n_labels.size(0), num_classes, n_labels.size(2),n_labels.size(3)).zero_() one_hot.scatter_(1,n_labels.data,1) return one_hot.cuda() ```