# DeepLearning-MachineLearning-Note

**Repository Path**: edata-code/DeepLearning-MachineLearning-Note

## Basic Information

- **Project Name**: DeepLearning-MachineLearning-Note
- **Description**: Deep Learning and machine learning note and code
- **Primary Language**: Python
- **License**: Apache-2.0
- **Default Branch**: master
- **Homepage**: https://www.cnblogs.com/endlesscoding/
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 1
- **Created**: 2020-01-06
- **Last Updated**: 2021-09-30

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

## 深度/机器学习笔记

深度学习和机器学习的一些笔记

### Focal Loss

- [ ] 添加更为详细的实现说明
  
FL对处理样本不平衡问题效果不错，这里对它做一个详细的解释，并且用pytorch实现，详细的解释请查看我的[博客](https://www.cnblogs.com/endlesscoding/p/12155588.html)。

- [x] Pytorch Focal Loss
- [x] Light Focal Loss
- [ ] Catboost Focal Loss

<img src="./FocalLoss/img/fl_paper_fig1.jpg" style="text-align:center"/>

### GHM Loss

GHM Loss 出自[Gradient Harmonized Single-Stage Detector](https://www.aaai.org/ojs/index.php/AAAI/article/view/4877/4750)，可以看作是Focal Loss的改进版，比Focal Loss的效果更好。

<img src='./GHM/img/GHM-fig1.jpg' style='text-align:center'/>

论文中实现了分类loss和回归loss。

#### 分类loss：GHM-C

$$
\begin{aligned}
L_{G H M-C} &=\frac{1}{N} \sum_{i=1}^{N}{\beta_{i} L_{C E}\left(p_{i}, p_{i}^{*}\right)} \\
&=\sum_{i=1}^{N}{\frac{L_{C E}\left(p_{i}, p_{i}^{*}\right)}{G D\left(g_{i}\right)}}
\end{aligned}
$$

#### 回归loss: GHM-R

提出一个`Smooth L1 Loss`替代版：

$$
A S L_{1}(d)=\sqrt{d^{2}+\mu^{2}}-\mu
$$

$$
\begin{aligned}
L_{G H M-R} 
&=\frac{1}{N} \sum_{i=1}^{N} \beta_{i} A S L_{1}\left(d_{i}\right) \\
&=\sum_{i=1}^{N} \frac{A S L_{1}\left(d_{i}\right)}{G D\left(g r_{i}\right)}
\end{aligned}
$$

> 更加详细的解释请看论文。