Label smooth loss 标签平滑损失函数 （附代码）

论文地址：https://proceedings.neurips.cc/paper_files/paper/2019/file/f1748d6b0fd9d439f71450117eba2725-Paper.pdf

1.是什么?

Label smooth loss是一种防止过拟合的正则化方法，它是在传统的分类loss采用softmax loss的基础上进行改进的。传统的softmax loss对于标签采用的是hard label的方式进行one hot编码，而Label smooth loss则对hard label得到的one hot编码添加一点点噪声，使得模型更加鲁棒。具体来说，Label smooth loss会将真实标签的概率值从1降低到一个较小的值，同时将其他标签的概率值从0提高到一个较小的值，这样可以使得模型更加关注数据的类别情况，从而提高模型的泛化能力。同时，Label smooth loss还可以减少模型对于噪声数据的过拟合，提高模型的鲁棒性。

NIPS 2019上的这篇论文When Does Label Smoothing Help?用实验说明了为什么Label smoothing可以work，指出标签平滑可以让分类之间的cluster更加紧凑，增加类间距离，减少类内距离，提高泛化性，同时还能提高Model Calibration（模型对于预测值的confidences和accuracies之间aligned的程度）。但是在模型蒸馏中使用Label smoothing会导致性能下降。
 

2.为什么？

 label smooth是相对于hard label和soft label 而言的，一般的分类任务中我们对label是采用hard label的方式进行one hot编码，而对hard label得到的one hot编码添加一点点噪声。举例如下图来自——如何理解soft target这一做法？：

hard label和soft label的优缺点在图中也给出来了，相对来说soft label拥有携带更多的信息，更好的描述数据的类别情况，而hard label丢失了类内和类间的关联，从这个角度来看soft label确实能在一定程度上提高模型的泛化能力，也就是相同数据能提点。

3.怎么样？

3.1交叉熵损失函数在多分类任务中存在的问题

多分类任务中，神经网络会输出一个当前数据对应于各个类别的置信度分数，将这些分数通过softmax进行归一化处理，最终会得到当前数据属于每个类别的概率。

然后计算交叉熵损失函数： 

训练神经网络时，最小化预测概率和标签真实概率之间的交叉熵，从而得到最优的预测概率分布。最优的预测概率分布是： 

神经网络会促使自身往正确标签和错误标签差值最大的方向学习，在训练数据较少，不足以表征所有的样本特征的情况下，会导致网络过拟合。

3.2 label smoothing（标签平滑）

label smoothing可以解决上述问题，这是一种正则化策略，主要是通过soft one-hot来加入噪声，减少了真实样本标签的类别在计算损失函数时的权重，最终起到抑制过拟合的效果。

 增加label smoothing后真实的概率分布有如下改变：

最优预测概率分布如下：

这里的α是任意实数，最终模型通过抑制正负样本输出差值，使得网络有更强的泛化能力。 

3.3 代码实现

pytorch实现

"""
标签平滑
可以把真实标签平滑集成在loss函数里面，然后计算loss
也可以直接在loss函数外面执行标签平滑，然后计算散度loss
"""
import torch.nn as nn
import torch
 
class LabelSmoothingLoss(nn.Module):
    """
    标签平滑Loss
    """
    def __init__(self, classes, smoothing=0.0, dim=-1):
        """
        :param classes: 类别数目
        :param smoothing: 平滑系数
        :param dim: loss计算平均值的维度
        """
        super(LabelSmoothingLoss, self).__init__()
        self.confidence = 1.0 - smoothing
        self.smoothing = smoothing
        self.cls = classes
        self.dim = dim
        self.loss = nn.KLDivLoss()
 
    def forward(self, pred, target):
        pred = pred.log_softmax(dim=self.dim)
        with torch.no_grad():
            # true_dist = pred.data.clone()
            true_dist = torch.zeros_like(pred)
            true_dist.fill_(self.smoothing / (self.cls - 1))
            true_dist.scatter_(1, target.data.unsqueeze(1), self.confidence)
        #torch.mean(torch.sum(-true_dist * pred, dim=self.dim))就是按照公式来计算损失
        loss = torch.mean(torch.sum(-true_dist * pred, dim=self.dim))
        #采用KLDivLoss来计算
        loss = self.loss(pred,true_dist)
        return loss

《Delving Deep into Label Smoothing》这篇论文就提供了一种在线标签平滑策略方法，使用一种在线学习的方式来生成soft label，相比传统的soft label方法，论文提出的方法声称效提高分类性能和模型的鲁棒性，优于LS、Bootsoft等方法。

import torch
import torch.nn as nn
from torch import Tensor
 
 
class OnlineLabelSmoothing(nn.Module):
    """
    Implements Online Label Smoothing from paper
    https://arxiv.org/pdf/2011.12562.pdf
    使用方法
    from ols import OnlineLabelSmoothing
    criterion = OnlineLabelSmoothing(alpha=..., n_classes=...)
    for epoch in range(...):  # loop over the dataset multiple times
        for i, data in enumerate(...):
            inputs, labels = data
            # zero the parameter gradients
            optimizer.zero_grad()
            # forward + backward + optimize
            outputs = net(inputs)
            loss = criterion(outputs, labels)
            loss.backward()
            optimizer.step()
        print(f'Epoch {epoch} finished!')
        # Update the soft labels for next epoch
        criterion.next_epoch()
        criterion.eval()
        dev()/test()
    """
 
    def __init__(self, alpha: float, n_classes: int, smoothing: float = 0.1):
        """
        :param alpha: Term for balancing soft_loss and hard_loss
        :param n_classes: Number of classes of the classification problem
        :param smoothing: Smoothing factor to be used during first epoch in soft_loss
        """
        super(OnlineLabelSmoothing, self).__init__()
        assert 0 <= alpha <= 1, 'Alpha must be in range [0, 1]'
        self.a = alpha
        self.n_classes = n_classes
 
        # Initialize soft labels with normal LS for first epoch
        self.register_buffer('supervise', torch.zeros(n_classes, n_classes))
        self.supervise.fill_(smoothing / (n_classes - 1))
        self.supervise.fill_diagonal_(1 - smoothing)
 
        # Update matrix is used to supervise next epoch
        self.register_buffer('update', torch.zeros_like(self.supervise))
        # For normalizing we need a count for each class
        self.register_buffer('idx_count', torch.zeros(n_classes))
        self.hard_loss = nn.CrossEntropyLoss()
 
    def forward(self, y_h: Tensor, y: Tensor):
        # Calculate the final loss
        soft_loss = self.soft_loss(y_h, y)
        hard_loss = self.hard_loss(y_h, y)
        return self.a * hard_loss + (1 - self.a) * soft_loss
 
    def soft_loss(self, y_h: Tensor, y: Tensor):
        """
        Calculates the soft loss and calls step
        to update `update`.
        :param y_h: Predicted logits.
        :param y: Ground truth labels.
        :return: Calculates the soft loss based on current supervise matrix.
        """
        y_h = y_h.log_softmax(dim=-1)
        if self.training:
            with torch.no_grad():
                self.step(y_h.exp(), y)
        true_dist = torch.index_select(self.supervise, 1, y).swapaxes(-1, -2)
        return torch.mean(torch.sum(-true_dist * y_h, dim=-1))
 
    def step(self, y_h: Tensor, y: Tensor) -> None:
        """
        Updates `update` with the probabilities
        of the correct predictions and updates `idx_count` counter for
        later normalization.
        Steps:
            1. Calculate correct classified examples.
            2. Filter `y_h` based on the correct classified.
            3. Add `y_h_f` rows to the `j` (based on y_h_idx) column of `memory`.
            4. Keep count of # samples added for each `y_h_idx` column.
            5. Average memory by dividing column-wise by result of step (4).
        Note on (5): This is done outside this function since we only need to
                     normalize at the end of the epoch.
        """
        # 1. Calculate predicted classes
        y_h_idx = y_h.argmax(dim=-1)
        # 2. Filter only correct
        mask = torch.eq(y_h_idx, y)
        y_h_c = y_h[mask]
        y_h_idx_c = y_h_idx[mask]
        # 3. Add y_h probabilities rows as columns to `memory`
        self.update.index_add_(1, y_h_idx_c, y_h_c.swapaxes(-1, -2))
        # 4. Update `idx_count`
        self.idx_count.index_add_(0, y_h_idx_c, torch.ones_like(y_h_idx_c, dtype=torch.float32))
 
    def next_epoch(self) -> None:
        """
        This function should be called at the end of the epoch.
        It basically sets the `supervise` matrix to be the `update`
        and re-initializes to zero this last matrix and `idx_count`.
        """
        # 5. Divide memory by `idx_count` to obtain average (column-wise)
        self.idx_count[torch.eq(self.idx_count, 0)] = 1  # Avoid 0 denominator
        # Normalize by taking the average
        self.update /= self.idx_count
        self.idx_count.zero_()
        self.supervise = self.update
        self.update = self.update.clone().zero_()

参考：

label smoothing(标签平滑)学习笔记

label smooth标签平滑的理解