README.md 5.1 KB
Newer Older
dcuai's avatar
dcuai committed
1
# MobileNetV2
sunxx1's avatar
sunxx1 committed
2

sunxx1's avatar
sunxx1 committed
3
4
5
6
7
8
## 论文

MobileNetV2: Inverted Residuals and Linear Bottlenecks

- https://openaccess.thecvf.com/content_cvpr_2018/papers/Sandler_MobileNetV2_Inverted_Residuals_CVPR_2018_paper.pdf

sunxx1's avatar
sunxx1 committed
9
## 模型结构
sunxx1's avatar
sunxx1 committed
10
11
12

MobileNetV2是一种轻量级的卷积神经网络模型,由Google在2018年提出。它是MobileNet系列中的第二个版本,主要用于移动设备和嵌入式设备等资源受限的环境中进行图像分类、目标检测等计算机视觉任务。

renzhc's avatar
renzhc committed
13
![loading-ag-535](./images/d15a0e56517b4f7284a862f1d6eaef9a.png)
sunxx1's avatar
sunxx1 committed
14
15
16



sunxx1's avatar
sunxx1 committed
17
## 算法原理
sunxx1's avatar
sunxx1 committed
18

sunxx1's avatar
sunxx1 committed
19
MobileNetV2的网络结构主要由两部分组成:特征提取层和分类器。
sunxx1's avatar
sunxx1 committed
20

sunxx1's avatar
sunxx1 committed
21
![20231124104337](./images/20231124104337.png)
sunxx1's avatar
sunxx1 committed
22

sunxx1's avatar
sunxx1 committed
23
## 环境配置
sunxx1's avatar
sunxx1 committed
24

sunxx1's avatar
sunxx1 committed
25
### Docker(方法一)
sunxx1's avatar
sunxx1 committed
26

renzhc's avatar
renzhc committed
27
28
29
推荐使用docker方式运行,拉取提供的docker镜像

```shell
renzhc's avatar
renzhc committed
30
docker pull image.sourcefind.cn:5000/dcu/admin/base/pytorch:2.1.0-ubuntu20.04-dtk24.04.1-py3.10
renzhc's avatar
renzhc committed
31
32
33
34
35
36
37
38
39
```

基于拉取的镜像创建容器

```shell
# <your IMAGE ID or NAME>用以上拉取的docker的镜像ID或名称替换
docker run -it --name=mobilenetv2 --network=host --ipc=host --shm-size=16g  --device=/dev/kfd --device=/dev/dri --device=/dev/mkfd --group-add video --privileged --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -v /opt/hyhal:/opt/hyhal:ro -v $PWD/mobilenetv2_mmcv:/home/mobilenetv2_mmcv <your IMAGE ID> bash
```

renzhc's avatar
renzhc committed
40
克隆并安装git仓库,安装相关依赖
sunxx1's avatar
sunxx1 committed
41

renzhc's avatar
renzhc committed
42
```python
chenzk's avatar
chenzk committed
43
git clone --recursive http://developer.sourcefind.cn/codes/modelzoo/mobilenetv2_mmcv.git
renzhc's avatar
renzhc committed
44
cd mobilenetv2_mmcv/mmpretrain-mmcv
renzhc's avatar
renzhc committed
45
pip install -e .
sunxx1's avatar
sunxx1 committed
46
47
48
49
50
pip install -r requirements.txt
```

### Dockerfile(方法二)

renzhc's avatar
renzhc committed
51
```shell
sunxx1's avatar
sunxx1 committed
52
53
cd mobilenetv2_mmcv/docker
docker build --no-cache -t mobilenetv2_mmcv:latest .
renzhc's avatar
renzhc committed
54
docker run -it --name=mobilenetv2 --network=host --ipc=host --shm-size=16g  --device=/dev/kfd --device=/dev/dri --device=/dev/mkfd --group-add video --privileged --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -v /opt/hyhal:/opt/hyhal:ro -v $PWD/mobilenetv2_mmcv:/home/mobilenetv2_mmcv <your IMAGE ID> bash
renzhc's avatar
renzhc committed
55
56
57
pip install -e .
# 若遇到Dockerfile启动的方式安装环境需要长时间等待,可注释掉里面的pip安装,启动容器后再安装python库:
pip install -r requirements.txt
sunxx1's avatar
sunxx1 committed
58
59
60
61
```

### Anaconda(方法三)

chenzk's avatar
chenzk committed
62
1、关于本项目DCU显卡所需的特殊深度学习库可从光合开发者社区下载安装: https://developer.sourcefind.cn/tool/
sunxx1's avatar
sunxx1 committed
63
64

```plaintext
renzhc's avatar
renzhc committed
65
66
67
68
69
DTK驱动: DTK-24.04.1
python==3.10
torch==2.1.0
torchvision==0.16.0+das1.1.git7d45932.abi1.dtk2404.torch2.1
mmcv==2.0.1+das1.1.gite58da25.abi1.dtk2404.torch2.1.0
sunxx1's avatar
sunxx1 committed
70
71
72
Tips:以上dtk驱动、python、torch等DCU相关工具版本需要严格一一对应
```

renzhc's avatar
renzhc committed
73
74
75
76
77
78
79
80
2、安装mmpretrain仓库源码

```shell
cd mobilenetv2_mmcv/mmpretrain-mmcv
pip install -e .
```

3、其它非特殊库参照requirements.txt安装
sunxx1's avatar
sunxx1 committed
81

renzhc's avatar
renzhc committed
82
```shell
sunxx1's avatar
sunxx1 committed
83
84
pip install -r requirements.txt
```
sunxx1's avatar
sunxx1 committed
85

sunxx1's avatar
sunxx1 committed
86
## 数据集
sunxx1's avatar
sunxx1 committed
87

88
89
### ImageNet

chenzk's avatar
chenzk committed
90
在本项目中可以使用ImageNet数据集。下载imagenet-2012数据集:https://image-net.org/
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124

下载其中的ILSVRC2012_img_train.tar和ILSVRC2012_img_val.tar并按照以下方式解包

```bash
cd mmpretrain-mmcv/data/imagenet
mkdir train && cd train
tar -xvf ILSVRC2012_img_train.tar
```

解包后是1000个tar文件,每个tar对应了一个类别,解包至对应文件夹

```bash
for tarfile in *.tar; do
    dirname="${tarfile%.tar}"
    mkdir "$dirname"
    tar -xvf "$tarfile" -C "$dirname"
done
```

目录结构如下

```
data
└── imagenet
    ├── train
    │   ├── n01440764
    │   │   ├── n01440764_10026.JPEG
    │   │   ├── n01440764_10027.JPEG
    ├──val
    │   ├── n01440764 
    │   │   ├── ILSVRC2012_val_00000293.JPEG
```


renzhc's avatar
renzhc committed
125

renzhc's avatar
renzhc committed
126
### Tiny-ImageNet-200
renzhc's avatar
renzhc committed
127

chenzk's avatar
chenzk committed
128
由于ImageNet完整数据集较大,可以使用[tiny-imagenet-200](http://cs231n.stanford.edu/tiny-imagenet-200.zip)进行测试,此时需要对配置脚本进行一些修改,可参照mmpretrain-mmcv子仓库进行配置,其中提供了使用Tiny-ImageNet-200进行训练的若干配置脚本。
sunxx1's avatar
sunxx1 committed
129

renzhc's avatar
renzhc committed
130
将训练数据集解压后放置于mmpretrain-mmcv/data/,对于Tiny-ImageNet,目录结构如下:
sunxx1's avatar
sunxx1 committed
131

sunxx1's avatar
sunxx1 committed
132
```
dcuai's avatar
dcuai committed
133
data
renzhc's avatar
renzhc committed
134
135
136
137
138
139
└── imagenet
    ├── test/
    ├── train/
    ├── val/
    ├── wnids.txt
    └── words.txt
sunxx1's avatar
sunxx1 committed
140
```
renzhc's avatar
renzhc committed
141

renzhc's avatar
renzhc committed
142
## 训练
renzhc's avatar
renzhc committed
143

renzhc's avatar
renzhc committed
144
Tiny-ImageNet-200
sunxx1's avatar
sunxx1 committed
145

renzhc's avatar
renzhc committed
146
147
148
```shell
bash tools/dist_train.sh mobilenet-v2-test.py 8
```
sunxx1's avatar
sunxx1 committed
149

renzhc's avatar
renzhc committed
150
ImageNet
sunxx1's avatar
sunxx1 committed
151

renzhc's avatar
renzhc committed
152
153
154
```shell
bash tools/dist_train.sh configs/mobilenet_v2/mobilenet-v2_8xb32_in1k.py 8
```
sunxx1's avatar
sunxx1 committed
155

renzhc's avatar
renzhc committed
156
157
158
159
tips:如需其他卡数训练,将命令中的8改为所需卡数即可;如遇端口占用问题,可在tools/dist_train.sh修改端口。

## Result

chenzk's avatar
chenzk committed
160
![img](https://developer.sourcefind.cn/codes/modelzoo/vit_pytorch/-/raw/master/image/README/1695381570003.png)
renzhc's avatar
renzhc committed
161
162
163
164

### 精度

测试数据使用的是ImageNet数据集,使用的加速卡是DCU Z100L。
sunxx1's avatar
sunxx1 committed
165

renzhc's avatar
renzhc committed
166
167
168
| 卡数  | 精度                        |
|:---:|:-------------------------:|
| 8   | top1:0.71764;top5:0.90386 |
sunxx1's avatar
sunxx1 committed
169

sunxx1's avatar
sunxx1 committed
170
171
172
173
174
175
176
177
## 应用场景

### 算法类别

图像分类

### 热点行业

178
制造,能源,交通,网安,安防
sunxx1's avatar
sunxx1 committed
179

dcuai's avatar
dcuai committed
180
## 源码仓库及问题反馈
sunxx1's avatar
sunxx1 committed
181

chenzk's avatar
chenzk committed
182
https://developer.sourcefind.cn/codes/modelzoo/mobilenetv2_mmcv
sunxx1's avatar
sunxx1 committed
183

dcuai's avatar
dcuai committed
184
## 参考资料
sunxx1's avatar
sunxx1 committed
185

dcuai's avatar
dcuai committed
186
https://github.com/open-mmlab/mmpretrain