解决样本不均衡问题

重采样

 label_id_level_1 = int(example.label_level_1)
 label_id_level_2 = int(example.label_level_2)
 samp_weight = math.sqrt(1 /label2freq_level_2[label_list_level_2[label_id_level_2]])
 sample_weights.append(samp_weight)

Pytorch中使用weightedRandomSampler进行样本出现概率的设置，提升小类出现的概率

 def train(self):
        if self.args.use_weighted_sampler:
            train_sampler = WeightedRandomSampler(
                self.train_sample_weights,
                len(self.train_sample_weights),
            )
        else:
            train_sampler = RandomSampler(self.train_dataset)
        train_dataloader = DataLoader(
            self.train_dataset,
            sampler=train_sampler,
            batch_size=self.args.train_batch_size

重加权

 # class weights
        class_weights_level_1 = []
        for i, lab in enumerate(label_list_level_1):
            class_weights_level_1.append(label2freq_level_1[lab])
        class_weights_level_1 = [1/w for w in class_weights_level_1]
        if self.args.use_weighted_sampler:
            class_weights_level_1 = [math.sqrt(w) for w in class_weights_level_1]
        else:
            class_weights_level_1 = [w for w in class_weights_level_1]
        print("class_weights_level_1: ", class_weights_level_1)
        self.class_weights_level_1 = F.softmax(torch.FloatTensor(
            class_weights_level_1
        ).to(self.args.device))

 # 1. loss
        if label_ids_level_2 is not None:
            if self.args.use_focal_loss:
                loss_fct = FocalLoss(
                    self.num_labels_level_2,
                    alpha=self.class_weights_level_2,
                    gamma=self.args.focal_loss_gamma,
                    size_average=True
                )
            elif self.args.use_class_weights:
                loss_fct = nn.CrossEntropyLoss(weight=self.class_weights_level_2)
            else:
                loss_fct = nn.CrossEntropyLoss()

            loss_level_2 = loss_fct(
                logits_level_2.view(-1, self.num_labels_level_2),
                label_ids_level_2.view(-1)
            )
            outputs = (loss_level_2,) + outputs

        return outputs

Focal loss

class FocalLoss(nn.Module):
    r"""
        This criterion is a implemenation of Focal Loss, which is proposed in
        Focal Loss for Dense Object Detection.

            Loss(x, class) = - \alpha (1-softmax(x)[class])^gamma \log(softmax(x)[class])

        The losses are averaged across observations for each minibatch.

        Args:
            alpha(1D Tensor, Variable) : the scalar factor for this criterion
            gamma(float, double) : gamma > 0; reduces the relative loss for well-classified examples (p > .5),
                                   putting more focus on hard, misclassified examples
            size_average(bool): By default, the losses are averaged over observations for each minibatch.
                                However, if the field size_average is set to False, the losses are
                                instead summed for each minibatch.


    """

    def __init__(self, class_num, alpha=None, gamma=2, size_average=True, device=None):
        super(FocalLoss, self).__init__()
        if alpha is None:
            self.alpha = torch.ones(class_num, 1).to(device)
        else:
            self.alpha = alpha.to(device)

        self.gamma = gamma
        self.class_num = class_num
        self.size_average = size_average

    def forward(self, inputs, targets):
        N = inputs.size(0)
        C = inputs.size(1)
        P = F.softmax(inputs)

        class_mask = torch.zeros_like(inputs).to(inputs.device)
        ids = targets.view(-1, 1)
        class_mask.scatter_(1, ids.data, 1.)
        # print("class_mask: ", class_mask)

        if inputs.is_cuda and not self.alpha.is_cuda:
            self.alpha = self.alpha.cuda()
        alpha = self.alpha[ids.data.view(-1)]
        # print("alpha: ", alpha)

        probs = (P * class_mask).sum(1).view(-1, 1)
        # print("probs: ", probs)

        log_p = probs.log()
        # print('log_p size= {}'.format(log_p.size()))
        # print(log_p)

        batch_loss = -alpha * (torch.pow((1 - probs), self.gamma)) * log_p
        # print('-----bacth_loss------')
        # print(batch_loss)

        if self.size_average:
            loss = batch_loss.mean()
        else:
            loss = batch_loss.sum()
        return loss

 from typing import List, Union

import torch
import torch.nn as nn
import torch.nn.functional as F


class FocalLoss(nn.Module):
    """
    Focal loss(https://arxiv.org/pdf/1708.02002.pdf)
    Shape:
        - input: (N, C)
        - target: (N)
        - Output: Scalar loss
    Examples:
        >>> loss = FocalLoss(gamma=2, alpha=[1.0]*7)
        >>> input = torch.randn(3, 7, requires_grad=True)
        >>> target = torch.empty(3, dtype=torch.long).random_(7)
        >>> output = loss(input, target)
        >>> output.backward()
    """
    def __init__(self, gamma=0, alpha=None, reduction="none"):
        super(FocalLoss, self).__init__()
        self.gamma = gamma
        self.alpha = alpha
        if alpha is not None:
            if isinstance(alpha, list):
                self.alpha = torch.FloatTensor(alpha)
            else:
                self.alpha = alpha

        self.reduction = reduction

    def forward(self, input, target):
        '''
        - input: (N, C), logits
        - target: (N)
        - Output: Scalar loss

        Parameters
        ----------
        input
        target

        Returns
        -------

        '''

        # [N, 1]
        target = target.unsqueeze(-1)
        # [N, C]
        pt = F.softmax(input, dim=-1)
        logpt = F.log_softmax(input, dim=-1)

        # 得到答案标签所获得 概率值 和 对数概率值
        # [N]
        pt = pt.gather(1, target).squeeze(-1)
        logpt = logpt.gather(1, target).squeeze(-1)

        # 加上 class weights
        if self.alpha is not None:
            # [N] at[i] = alpha[target[i]]
            # 得到每个样本应该得到的 class weights
            at = self.alpha.gather(0, target.squeeze(-1))
            logpt = logpt * at.to(logpt.device)

        loss = -1 * (1 - pt) ** self.gamma * logpt
        if self.reduction == "none":
            return loss
        if self.reduction == "mean":
            return loss.mean()
        return loss.sum()

    @staticmethod
    def convert_binary_pred_to_two_dimension(x, is_logits=True):
        """
        Args:
            x: (*): (log) prob of some instance has label 1
            is_logits: if True, x represents log prob; otherwhise presents prob
        Returns:
            y: (*, 2), where y[*, 1] == log prob of some instance has label 0,
                             y[*, 0] = log prob of some instance has label 1
        """
        probs = torch.sigmoid(x) if is_logits else x
        probs = probs.unsqueeze(-1)
        probs = torch.cat([1-probs, probs], dim=-1)
        logprob = torch.log(probs+1e-4)  # 1e-4 to prevent being rounded to 0 in fp16
        return logprob

    def __str__(self):
        return f"Focal Loss gamma:{self.gamma}"

    def __repr__(self):
        return str(self)

Dice Loss

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import Tensor
from typing import Optional


class DiceLoss(nn.Module):
    """
    Dice coefficient for short, is an F1-oriented statistic
    used to gauge the similarity of two sets.
    Given two sets A and B, the vanilla dice coefficient
    between them is given as follows:
        Dice(A, B)  = 2 * True_Positive / (2 * True_Positive + False_Positive + False_Negative)
                    = 2 * |A and B| / (|A| + |B|)

    Math Function:
        U-NET: https://arxiv.org/abs/1505.04597.pdf
        dice_loss(p, y) = 1 - numerator / denominator
            numerator = 2 * \sum_{1}^{t} p_i * y_i + smooth
            denominator = \sum_{1}^{t} p_i + \sum_{1} ^{t} y_i + smooth
        if square_denominator is True, the denominator is
            \sum_{1}^{t} (p_i ** 2) + \sum_{1} ^{t} (y_i ** 2) + smooth
        V-NET: https://arxiv.org/abs/1606.04797.pdf
    Args:
        smooth (float, optional): a manual smooth value for numerator and denominator.
        square_denominator (bool, optional): [True, False], specifies whether to square the denominator in the loss function.
        with_logits (bool, optional): [True, False], specifies whether the
                input tensor is normalized by Sigmoid/Softmax funcs.
        ohem_ratio: OHEM（online hard example miniing）
                max ratio of positive/negative, defautls to 0.0, which means no ohem.
        alpha: dsc alpha
    Shape:
        - input: (*)
        - target: (*)
        - mask: (*) 0,1 mask for the input sequence.
        - Output: Scalar loss
    Examples:
        >>> loss = DiceLoss(with_logits=True, ohem_ratio=0.1)
        >>> input = torch.FloatTensor([2, 1, 2, 2, 1])
        >>> input.requires_grad=True
        >>> target = torch.LongTensor([0, 1, 0, 0, 0])
        >>> output = loss(input, target)
        >>> output.backward()
    """
    def __init__(self,
                 smooth: Optional[float] = 1e-4,
                 square_denominator: Optional[bool] = False,
                 with_logits: Optional[bool] = True,
                 ohem_ratio: float = 0.0,
                 alpha: float = 0.0,
                 reduction: Optional[str] = "mean",
                 index_label_position=True) -> None:
        super(DiceLoss, self).__init__()

        self.reduction = reduction
        self.with_logits = with_logits
        self.smooth = smooth
        self.square_denominator = square_denominator
        self.ohem_ratio = ohem_ratio
        self.alpha = alpha
        self.index_label_position = index_label_position

    def forward(self, input: Tensor, target: Tensor, mask: Optional[Tensor] = None) -> Tensor:
        logits_size = input.shape[-1]

        if logits_size != 1:
            loss = self._multiple_class(input, target, logits_size, mask=mask)
        else:
            loss = self._binary_class(input, target, mask=mask)

        if self.reduction == "mean":
            return loss.mean()
        if self.reduction == "sum":
            return loss.sum()
        return loss

    def _compute_dice_loss(self, flat_input, flat_target):
        flat_input = ((1 - flat_input) ** self.alpha) * flat_input
        interection = torch.sum(flat_input * flat_target, -1)
        if not self.square_denominator:
            loss = 1 - ((2 * interection + self.smooth) /
                        (flat_input.sum() + flat_target.sum() + self.smooth))
        else:
            loss = 1 - ((2 * interection + self.smooth) /
                        (torch.sum(torch.square(flat_input, ), -1) + torch.sum(torch.square(flat_target), -1) + self.smooth))

        return loss

    def _multiple_class(self, input, target, logits_size, mask=None):
        # input: [N, C]
        flat_input = input
        flat_input = torch.nn.Softmax(dim=1)(flat_input) if self.with_logits else flat_input

        # [N, ] --> [N, C]
        flat_target = F.one_hot(target, num_classes=logits_size).float() \
            if self.index_label_position else target.float()

        if mask is not None:
            mask = mask.float()
            flat_input = flat_input * mask
            flat_target = flat_target * mask
        else:
            mask = torch.ones_like(target)

        loss = None
        if self.ohem_ratio > 0:
            mask_neg = torch.logical_not(mask)
            for label_idx in range(logits_size):
                # logits_size: 类别数
                # pos_example： 标签为 label_idx的是正例；
                # neg_example： 标签不为 label_idx的是负例；
                pos_example = target == label_idx
                neg_example = target != label_idx

                pos_num = pos_example.sum()
                neg_num = mask.sum() - (pos_num - (mask_neg & pos_example).sum())
                keep_num = min(int(pos_num * self.ohem_ratio / logits_size), neg_num)

                if keep_num > 0:
                    # masked_select： 返回一个1-D tensor，根据flat_input对应于mask=1的部分的值返回
                    # 得到负样本对于本标签的打分
                    neg_scores = torch.masked_select(
                        flat_input,
                        neg_example.view(-1, 1).bool()
                    ).view(-1, logits_size)
                    neg_scores_idx = neg_scores[:, label_idx]
                    neg_scores_sort, _ = torch.sort(neg_scores_idx, )
                    threshold = neg_scores_sort[-keep_num + 1]

                    # 预测为本标签或者正确标签是本标签
                    cond = (torch.argmax(flat_input, dim=1) == label_idx & flat_input[:, label_idx] >= threshold) | pos_example.view(-1)
                    ohem_mask_idx = torch.where(cond, 1, 0)

                    flat_input_idx = flat_input[:, label_idx]
                    flat_target_idx = flat_target[:, label_idx]

                    flat_input_idx = flat_input_idx * ohem_mask_idx
                    flat_target_idx = flat_target_idx * ohem_mask_idx
                else:
                    flat_input_idx = flat_input[:, label_idx]
                    flat_target_idx = flat_target[:, label_idx]

                loss_idx = self._compute_dice_loss(
                    flat_input_idx.view(-1, 1),
                    flat_target_idx.view(-1, 1)
                )
                if loss is None:
                    loss = loss_idx
                else:
                    loss += loss_idx
            return loss

        else:
            for label_idx in range(logits_size):
                pos_example = target == label_idx
                flat_input_idx = flat_input[:, label_idx]  # 概率值
                flat_target_idx = flat_target[:, label_idx]  # 有正样本，有负样本 【N, 】

                loss_idx = self._compute_dice_loss(flat_input_idx.view(-1, 1), flat_target_idx.view(-1, 1))
                if loss is None:
                    loss = loss_idx
                else:
                    loss += loss_idx
            return loss

    def _binary_class(self, input, target, mask=None):
        flat_input = input.view(-1)
        flat_target = target.view(-1).float()
        flat_input = torch.sigmoid(flat_input) if self.with_logits else flat_input

        if mask is not None:
            mask = mask.float()
            flat_input = flat_input * mask
            flat_target = flat_target * mask
        else:
            mask = torch.ones_like(target)

        if self.ohem_ratio > 0:
            pos_example = target > 0.5
            neg_example = target <= 0.5
            mask_neg_num = mask <= 0.5

            pos_num = pos_example.sum() - (pos_example & mask_neg_num).sum()
            neg_num = neg_example.sum()
            keep_num = min(int(pos_num * self.ohem_ratio), neg_num)

            neg_scores = torch.masked_select(flat_input, neg_example.bool())
            neg_scores_sort, _ = torch.sort(neg_scores, )
            threshold = neg_scores_sort[-keep_num+1]
            cond = (flat_input > threshold) | pos_example.view(-1)
            ohem_mask = torch.where(cond, 1, 0)
            flat_input = flat_input * ohem_mask
            flat_target = flat_target * ohem_mask

        return self._compute_dice_loss(flat_input, flat_target)

    def __str__(self):
        return f"Dice Loss smooth:{self.smooth}, ohem: {self.ohem_ratio}, alpha: {self.alpha}"

    def __repr__(self):
        return str(self)