Mxnet Focal Loss实现

这个ssd有很多loss：

https://github.com/wshuail/SSD.Gluon/blob/a291f9562f5cc251d72cc1939c838a8bdc586d2d/loss.py

https://github.com/wshuail/gluon-detector/blob/2f7723411e7aff31dea7e632f9fe3b190ca5c294/lib/cus_loss/retinanet.py

import mxnet as mx
import numpy as npclass FocalLoss(mx.metric.EvalMetric):def __init__(self, num=None):super(FocalLoss, self).__init__('focalloss', num)self.eps = 1e-12self.gamma = 2self.alpha = 0.25def update(self, labels, preds):for label, pred in zip(labels, preds):label = label.asnumpy()pred = pred.asnumpy()label = label.ravel()assert label.shape[0] == pred.shape[0]prob = pred[np.arange(label.shape[0]), np.int64(label)]self.sum_metric += self.alpha * (-1.0 * np.power(1.0-prob, self.gamma) * np.log(prob + self.eps)).sum()self.num_inst += label.shape[0]

from focal_loss_OptimizedVersion import *
label = mx.sym.Variable('focalloss_label')
net = mx.symbol.Custom(data=net, op_type='FocalLoss', labels = label, name='focalloss', alpha=0.25, gamma=2)

from metric import *
eval_metric = mx.metric.CompositeEvalMetric()
eval_metric.add(FocalLoss())model = mx.mod.Module(context=mx.gpu(0),symbol=symbol,label_names=('focalloss_label',))model.fit(...,eval_metric=eval_metric,...)

# -*- coding: utf-8 -*-
from mxnet.gluon import loss
from mxnet import ndclass FocalLoss(loss.Loss):def __init__(self, gamma, batch_axis=0):super(FocalLoss, self).__init__(batch_axis=batch_axis, weight=None)self.gamma = gammadef hybrid_forward(self, F, predict, label, *args, **kwargs):assert label.shape == predict.shapelog_predict = F.log_softmax(predict, axis=1) * label  #提取前景部分predict_act = F.exp(log_predict)focal_loss = (-1 * (1 - predict_act) ** self.gamma * log_predict).mean(axis=(2, 3, 4))return F.mean(focal_loss, self._batch_axis)def __repr__(self):return self.__class__.__name__原文链接：https://blog.csdn.net/weixin_42071277/article/details/114632539

Mxnet Focal Loss实现相关推荐

Focal Loss
转自:https://blog.csdn.net/u014380165/article/details/77019084 论文:Focal Loss for Dense Object Detectio ...
一、Focal Loss理论及代码实现
文章目录前言一.基本理论二.实现 1.公式 2.代码实现 1.基于二分类交叉熵实现. 2.知乎大佬的实现前言本文参考:几时见得清梦博主文章参考原文:https://www.jianshu. ...
剖析Focal Loss损失函数：消除类别不平衡+挖掘难分样本 | CSDN博文精选
作者 | 图像所浩南哥来源 | CSDN博客论文名称:< Focal Loss for Dense Object Detection > 论文下载:https://arxiv.org/ ...
NeurIPS 2020 | Focal Loss改进版来了！GFocal Loss：良心技术，无Cost涨点！
本文作者:李翔 https://zhuanlan.zhihu.com/p/147691786 本文仅供学习参考,如有侵权,请联系删除! 论文地址:https://arxiv.org/abs/2006. ...
清晰易懂的Focal Loss原理解释
来源:https://www.cnblogs.com/king-lps/p/9497836.html 编辑:石头 1. 总述 Focal loss主要是为了解决one-stage目标检测中正负样本比例 ...
无痛涨点！大白话讲解 Generalized Focal Loss
点击上方"小白学视觉",选择加"星标"或"置顶" 重磅干货,第一时间送达作者丨李翔来源丨https://zhuanlan.zhihu.c ...
目标检测--Focal Loss for Dense Object Detection
Focal Loss for Dense Object Detection ICCV2017 https://arxiv.org/abs/1708.02002 本文算是用简单的方法解决复杂的问题了,好 ...
堪比Focal Loss！解决目标检测中样本不平衡的无采样方法
训练目标检测模型的一个难点是样本不均衡,特别是正负样本比例严重失衡.目前解决这类问题主要是两种方案(见综述Imbalance Problems in Object Detection: A Revie ...
Focal Loss和它背后的男人RetinaNet
说起Focal Loss,相信做CV的都不会陌生,当面临正负样本不平衡时可能第一个想到的就是用Focal Loss试试.但是怕是很多人会不知道这篇论文中所提出的one stage目标检测模型Retin ...

Mxnet Focal Loss实现

Mxnet Focal Loss实现相关推荐

最新文章

热门文章