pytorch版本的多分类focal loss

---

正文

代码参照某博主大佬写的。使用时有报错就自己改了改。

 logits = logits[..., None]
 labels = labels[..., None]

因为我的网络输出是二维的，label是一维的。所以上述改了一下。

device=logits.device

因为训练时报错说loss数据不在同一个空间，所以上面把device设置一下。全部代码如下：

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


class FocalLoss(nn.Module):
    def __init__(self, gamma=2, alpha=1, size_average=True):
        super(FocalLoss, self).__init__()
        self.gamma = gamma
        self.alpha = alpha
        self.size_average = size_average
        self.elipson = 0.000001

    def forward(self, logits, labels):
        """
        cal culates loss
        logits: batch_size * labels_length * seq_length
        labels: batch_size * seq_length
      """
        logits = logits[..., None]
        labels = labels[..., None]
        if labels.dim() > 2:
            labels = labels.contiguous().view(labels.size(0), labels.size(1), -1)
            labels = labels.transpose(1, 2)
            labels = labels.contiguous().view(-1, labels.size(2)).squeeze()
        if logits.dim() > 3:
            logits = logits.contiguous().view(logits.size(0), logits.size(1), logits.size(2), -1)
            logits = logits.transpose(2, 3)
            logits = logits.contiguous().view(-1, logits.size(1), logits.size(3)).squeeze()
        assert (logits.size(0) == labels.size(0))
        assert (logits.size(2) == labels.size(1))
        batch_size = logits.size(0)
        labels_length = logits.size(1)
        seq_length = logits.size(2)

        # transpose labels into labels onehot
        new_label = labels.unsqueeze(1)
        label_onehot = torch.zeros([batch_size, labels_length, seq_length], device=logits.device).scatter_(1, new_label, 1)

        # calculate log
        log_p = F.log_softmax(logits)
        pt = label_onehot * log_p
        sub_pt = 1 - pt
        fl = -self.alpha * (sub_pt) ** self.gamma * log_p
        if self.size_average:
            return fl.mean()
        else:
            return fl.sum()

---