pytorch实现Focal Loss系列损失函数(针对SOT中的定位问题)，包括Focal Loss、GFocal Loss、VFocal Loss

---

1.Focal Loss

具体细节不在多说，这里只给出损失函数代码

论文：Focal Loss for Dense Object Detection

import torch
import torch.nn as nn
import torch.nn.functional as F
class FocalLoss(nn.Module):
    def __init__(self,alpha=0.25, gamma=2.0,use_sigmoid=True):
        super().__init__()
        self.alpha = alpha
        self.gamma = gamma
        self.use_sigmoid = use_sigmoid
        if use_sigmoid:
            self.sigmoid = nn.Sigmoid()

    def forward(self, pred: torch.Tensor, target: torch.Tensor):
        r"""
        Focal loss
        :param pred: shape=(B,  HW)
        :param label: shape=(B, HW)
        """
        if self.use_sigmoid:
            pred = self.sigmoid(pred)
        pred = pred.view(-1)
        label = target.view(-1)
        pos = torch.nonzero(label > 0).squeeze(1)
        pos_num = max(pos.numel(),1.0)
        mask = ~(label == -1)
        pred = pred[mask]
        label= label[mask]
        focal_weight = self.alpha *(label- pred).abs().pow(self.gamma) * (label> 0.0).float() + (1 - self.alpha) * pred.abs().pow(self.gamma) * (label<= 0.0).float()
        loss = F.binary_cross_entropy(pred, label, reduction='none') * focal_weight
        return loss.sum()/pos_num

2.GFocal Loss

论文:Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection

import torch
import torch.nn as nn
import torch.nn.functional as F
class GFocalLoss(nn.Module):
    def __init__(self, beta=2.0,use_sigmoid=True):
        super().__init__()
        self.beta = beta
        self.use_sigmoid = use_sigmoid
        if use_sigmoid:
            self.sigmoid = nn.Sigmoid()

    def forward(self, pred: torch.Tensor, target: torch.Tensor):
        r"""
        Focal loss
        :param pred: shape=(B,  HW)
        :param label: shape=(B, HW)
        """
        if self.use_sigmoid:
            pred = self.sigmoid(pred)
        pred = pred.view(-1)
        label = target.view(-1)
        pos = torch.nonzero(label > 0).squeeze(1)
        pos_num = max(pos.numel(),1.0)
        mask = ~(label == -1)
        pred = pred[mask]
        label= label[mask]
        scale_factor = (pred - label).abs().pow(self.beta)
        loss = F.binary_cross_entropy(pred, label, reduction='none') * scale_factor
        return loss.sum()/pos_num 

3.VFocal Loss

论文：VarifocalNet: An IoU-aware Dense Object Detector

import torch
import torch.nn as nn
import torch.nn.functional as F
class VFocalLoss(nn.Module):
    def __init__(self,alpha=0.75, gamma=2.0,use_sigmoid=True):
        super().__init__()
        self.alpha = alpha
        self.gamma = gamma
        self.use_sigmoid = use_sigmoid
        if use_sigmoid:
            self.sigmoid = nn.Sigmoid()

    def forward(self, pred: torch.Tensor, target: torch.Tensor):
        r"""
        Focal loss
        :param pred: shape=(B,  HW)
        :param label: shape=(B, HW)
        """
        if self.use_sigmoid:
            pred = self.sigmoid(pred)
        pred = pred.view(-1)
        label = target.view(-1)
        pos = torch.nonzero(label > 0).squeeze(1)
        pos_num = max(pos.numel(),1.0)
        mask = ~(label == -1)
        pred = pred[mask]
        label= label[mask]
        focal_weight = label * (label > 0.0).float() + self.alpha * (pred - label).abs().pow(self.gamma) * (label <= 0.0).float()
        loss = F.binary_cross_entropy(pred, label, reduction='none') * focal_weight
        return loss.sum()/pos_num