FocalLoss 用于 one-stage 目标检测算法(Retinanet),提升检测效果.
也可以被用于分类任务中,解决数据不平衡问题.
1. Github - DeepLabV3Plus-Pytorch
import torch
import torch.nn as nn
import torch.nn.functional as F
class FocalLoss(nn.Module):
def __init__(self, alpha=1, gamma=0, size_average=True, ignore_index=255):
super(FocalLoss, self).__init__()
self.alpha = alpha
self.gamma = gamma
self.ignore_index = ignore_index
self.size_average = size_average
def forward(self, inputs, targets):
ce_loss = F.cross_entropy(inputs, targets,
reduction='none', ignore_index=self.ignore_index)
pt = torch.exp(-ce_loss)
focal_loss = self.alpha * (1-pt)**self.gamma * ce_loss
if self.size_average:
return focal_loss.mean()
else:
return focal_loss.sum()
2. Github - Focal-Loss-Pytorch
Github - yatengLG/Focal-Loss-Pytorch
基于 Pytorch 的实现如:
#!-*- coding: utf-8 -*-
# @Author : LG
from torch import nn
import torch
from torch.nn import functional as F
class focal_loss(nn.Module):
def __init__(self, alpha=0.25, gamma=2, num_classes = 3, size_average=True):
"""
focal_loss损失函数, -α(1-yi)**γ *ce_loss(xi,yi)
步骤详细的实现了 focal_loss损失函数.
:param alpha: 阿尔法α,类别权重.
当α是列表时,为各类别权重;
当α为常数时,类别权重为[α, 1-α, 1-α, ....],
常用于目标检测算法中抑制背景类,
retainnet中设置为0.25
:param gamma: 伽马γ,难易样本调节参数. retainnet中设置为2
:param num_classes: 类别数量
:param size_average: 损失计算方式,默认取均值
"""
super(focal_loss,self).__init__()
self.size_average = size_average
if isinstance(alpha,list):
assert len(alpha)==num_classes
# α可以以list方式输入,
# size:[num_classes] 用于对不同类别精细地赋予权重
print("Focal_loss alpha = {}, 将对每一类权重进行精细化赋值".format(alpha))
self.alpha = torch.Tensor(alpha)
else:
assert alpha<1 #如果α为一个常数,则降低第一类的影响,在目标检测中为第一类
print("Focal_loss alpha = {} ,将对背景类进行衰减,请在目标检测任务中使用.".format(alpha))
self.alpha = torch.zeros(num_classes)
self.alpha[0] += alpha
self.alpha[1:] += (1-alpha) # α 最终为 [ α, 1-α, 1-α, 1-α, 1-α, ...] size:[num_classes]
self.gamma = gamma
def forward(self, preds, labels):
"""
focal_loss损失计算
:param preds: 预测类别. size:[B,N,C] or [B,C] 分
别对应与检测与分类任务, B 批次, N检测框数, C类别数
:param labels: 实际类别. size:[B,N] or [B]
:return:
"""
# assert preds.dim()==2 and labels.dim()==1
preds = preds.view(-1,preds.size(-1))
self.alpha = self.alpha.to(preds.device)
# 这里并没有直接使用log_softmax, 因为后面会用到softmax的结果(当然也可以使用log_softmax,然后进行exp操作)
preds_softmax = F.softmax(preds, dim=1)
preds_logsoft = torch.log(preds_softmax)
# 这部分实现nll_loss ( crossempty = log_softmax + nll )
preds_softmax = preds_softmax.gather(1,labels.view(-1,1))
preds_logsoft = preds_logsoft.gather(1,labels.view(-1,1))
self.alpha = self.alpha.gather(0,labels.view(-1))
loss = -torch.mul(torch.pow((1-preds_softmax), self.gamma), preds_logsoft)
# torch.pow((1-preds_softmax), self.gamma) 为focal loss中 (1-pt)**γ
loss = torch.mul(self.alpha, loss.t())
if self.size_average:
loss = loss.mean()
else:
loss = loss.sum()
return loss