README.md 4.3 KB
Newer Older
dongchy920's avatar
dongchy920 committed
1
2
3
4
5
6
7
8
9
10
11
12
# ArcFace
## 论文
- https://arxiv.org/pdf/1801.07698.pdf
## 模型结构
这篇文章提出一种新的用于人脸识别的损失函数:additive angular margin loss,直接在角度空间(angular space)中最大化分类界限,基于该损失函数训练得到人脸识别算法ArcFace。
<div align=center>
    <img src="./docs/arcface.png"/>
</div>

## 算法原理
通过训练深度卷积神经网络嵌入 (DCNN Embedding) 来进行人脸识别。  
ArcFace训练流程:  
dongchy920's avatar
dongchy920 committed
13
<div align=center>
dongchy920's avatar
dongchy920 committed
14
    <img src="./docs/legend.png"/>
dongchy920's avatar
dongchy920 committed
15
</div>  
dongchy920's avatar
dongchy920 committed
16
17
18
19
20
21
22
23
24

<div align=center>
    <img src="./docs/train.jpg"/>
</div>


## 环境配置
### Docker(方法一)
[光源](https://www.sourcefind.cn/#/service-list)中拉取docker镜像:
dongchy920's avatar
arcface  
dongchy920 committed
25
```
dongchy920's avatar
dongchy920 committed
26
docker pull image.sourcefind.cn:5000/dcu/admin/base/pytorch:1.13.1-centos7.6-dtk23.10-py310
dongchy920's avatar
arcface  
dongchy920 committed
27
```
dongchy920's avatar
dongchy920 committed
28
创建容器并挂载目录进行开发:
dongchy920's avatar
arcface  
dongchy920 committed
29
```
dongchy920's avatar
dongchy920 committed
30
31
32
33
34
docker run -it --name {name} --shm-size=1024G  --device=/dev/kfd --device=/dev/dri/ --privileged --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --ulimit memlock=-1:-1 --ipc=host --network host --group-add video -v /opt/hyhal:/opt/hyhal:ro -v {}:{} {docker_image} /bin/bash
# 修改1 {name} 需要改为自定义名称,建议命名{框架_dtk版本_使用者姓名},如果有特殊用途可在命名框架前添加命名
# 修改2 {docker_image} 需要需要创建容器的对应镜像名称,如: pytorch:1.10.0-centos7.6-dtk-23.04-py37-latest【镜像名称:tag名称】
# 修改3 -v 挂载路径到容器指定路径
pip install -r requirements.txt
dongchy920's avatar
arcface  
dongchy920 committed
35
```
dongchy920's avatar
dongchy920 committed
36
37
38
39
40
41
42
43
44
45
46
47
48
### Dockerfile(方法二)
```
cd docker
docker build --no-cache -t arcface_pytorch:1.0 .
docker run -it --name {name} --shm-size=1024G  --device=/dev/kfd --device=/dev/dri/ --privileged --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --ulimit memlock=-1:-1 --ipc=host --network host --group-add video -v /opt/hyhal:/opt/hyhal:ro -v {}:{} {docker_image} /bin/bash 
pip install -r requirements.txt
```
### Anaconda(方法三)
线上节点推荐使用conda进行环境配置。
创建python=3.10的conda环境并激活
```
conda create -n arcface python=3.10
conda activate arcface
dongchy920's avatar
arcface  
dongchy920 committed
49
50
```

dongchy920's avatar
dongchy920 committed
51
52
53
54
55
56
57
58
59
60
61
62
63
关于本项目DCU显卡所需的特殊深度学习库可从[光合](https://developer.hpccube.com/tool/)开发者社区下载安装。
```
DTK驱动:dtk23.10
python:python3.10
pytorch:1.13.1
torchvision:0.14.1
```
安装其他依赖包
```
pip install -r requirements.txt
```
## 数据集
`MS1MV2\IJBC`
dongchy920's avatar
arcface  
dongchy920 committed
64

dongchy920's avatar
dongchy920 committed
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
- 训练集[faces_emore.zip](https://pan.baidu.com/s/1S6LJZGdqcZRle1vlcMzHOQ)
下载后解压到当前目录  
数据目录结构如下:
```
 ── faces_emore
    |   agedb_30.bin
    |   calfw.bin
    |   cfp_ff.bin
    |   cfp_fp.bin
    |   cplfw.bin
    |   lfw.bin
    |   property
    |   train.idx
    |   train.rec
    |   vgg2_fp.bin
```
- 测试集[IJBC.zip](https://pan.baidu.com/s/1Ok4sqTO8vqAE_kG3zV1rqw?pwd=1234)  
解压分卷压缩文件:
```
# 将所有的分卷压缩文件放在一个文件夹中
zip -s 0 IJBC.zip --out IJBC_ALL.zip
dongchy920's avatar
dongchy920 committed
86
unzip IJBC_ALL.zip
dongchy920's avatar
dongchy920 committed
87
```
dongchy920's avatar
arcface  
dongchy920 committed
88

dongchy920's avatar
dongchy920 committed
89
90
## 训练
Backbone使用ResNet100,在MS1MV3数据集上的预训练权重文件为[model.pt](https://pan.baidu.com/s/1W-TisIZtZmRQz32hq5T6Uw?pwd=1234)  
dongchy920's avatar
arcface  
dongchy920 committed
91

dongchy920's avatar
dongchy920 committed
92
93
### 单机单卡
```
dongchy920's avatar
dongchy920 committed
94
python train_v2.py configs/ms1mv2_r100.py
dongchy920's avatar
dongchy920 committed
95
```
dongchy920's avatar
arcface  
dongchy920 committed
96

dongchy920's avatar
dongchy920 committed
97
98
### 单机多卡
```
dongchy920's avatar
dongchy920 committed
99
100
torchrun --nproc_per_node=4 train_v2.py configs/ms1mv2_r100.py
```
dongchy920's avatar
dongchy920 committed
101

102
### 测试
dongchy920's avatar
dongchy920 committed
103
104
105
106
下载权重文件和测试数据集,测试模型精度:
```
python eval_ijbc.py --model-prefix model.pt --image-path IJBC_ALL --network r100
```
dongchy920's avatar
dongchy920 committed
107
108
109
110
111
112

## result
<div align=center>
    <img src="./docs/ROC.png"/>
</div>

113
### 精度
dongchy920's avatar
dongchy920 committed
114
115
116
117
118
119
模型在MS1MV2数据集的测试指标:
| 模型 | 数据类型 | AUC |
| :------: | :------: | :------: |
| [r34](https://pan.baidu.com/s/1LR0zm8AxwN2tZH55xQdzHw?pwd=1234) | fp16 | 99.5611% |
| [r50](https://pan.baidu.com/s/128GP5J-jWvNbQAAur68bHw?pwd=1234) | fp16 | 99.5616% |
| [r100](https://pan.baidu.com/s/1e4Qg2i6wqyBCcwgA-dA8xw?pwd=1234) | fp16 | 99.5767% |
dongchy920's avatar
arcface  
dongchy920 committed
120

dongchy920's avatar
dongchy920 committed
121
122
123
## 应用场景
### 算法类别
人脸识别
dongchy920's avatar
arcface  
dongchy920 committed
124

dongchy920's avatar
dongchy920 committed
125
126
### 热点应用行业
安防,交通,教育
dongchy920's avatar
arcface  
dongchy920 committed
127
128


dongchy920's avatar
dongchy920 committed
129
130
131
132
## 源码仓库及问题反馈
[https://developer.hpccube.com/codes/modelzoo/arcface_pytorch](https://developer.hpccube.com/codes/modelzoo/arcface_pytorch)
## 参考资料
[https://github.com/deepinsight/insightface/tree/master/recognition/arcface_torch](https://github.com/deepinsight/insightface/tree/master/recognition/arcface_torch)
dongchy920's avatar
arcface  
dongchy920 committed
133
134