README.md 3.63 KB
Newer Older
zhanggezhong's avatar
zhanggezhong committed
1
2
3
4
5
# <div align="center"><strong>Mask-RCNN</strong></div>
## 简介
Mask R-CNN 是在 Faster R-CNN 架构的基础上发展而来的,Faster R-CNN 是一种经典的两阶段目标检测算法,主要由骨干网络(Backbone)、区域建议网络(Region Proposal Network, RPN)和检测头(Detection Head)三部分组成。Mask R-CNN 在 Faster R-CNN 的基础上添加了一个用于生成目标掩码(mask)的分支,形成了一个多任务学习框架。

## 安装
zhanggezhong's avatar
zhanggezhong committed
6
组件支持组合
zhanggezhong's avatar
zhanggezhong committed
7

zhanggezhong's avatar
zhanggezhong committed
8
9
10
11
12
   | PyTorch版本 | fastpt版本  |Mask-RCNN版本      | DTK版本                  | Python版本       | 推荐编译方式 |
   | ----------- | ----------- | ----------- | ------------------------ | -----------------| ------------ |
   | 2.5.1       | 2.1.0       |0.1        | >= 25.04                 | 3.8、3.10、3.11  | fastpt不转码 |
   | 2.4.1       | 2.0.1       |0.1        | >= 25.04                 | 3.8、3.10、3.11  | fastpt不转码 |
   
zhanggezhong's avatar
zhanggezhong committed
13

zhanggezhong's avatar
zhanggezhong committed
14
+ pytorch版本大于2.4.1 && dtk版本大于25.04 推荐使用fastpt不转码编译。
zhanggezhong's avatar
zhanggezhong committed
15

zhanggezhong's avatar
zhanggezhong committed
16
17
18
19
20
21
22
23
24
### 1、使用pip方式安装
maskrcnn whl包下载目录:[光和开发者社区](https://download.sourcefind.cn:65024/4/main),选择对应的pytorch版本和python版本下载对应maskrcnn的whl包
```shell
pip install torch* (下载torch的whl包)
pip install fastpt* --no-deps (下载fastpt的whl包)
source  /usr/local/bin/fastpt -E
pip install maskrcnn* (下载的maskrcnn-fastpt的whl包)
```
### 2、使用源码编译方式安装
zhanggezhong's avatar
zhanggezhong committed
25
26

#### 编译环境准备
zhanggezhong's avatar
zhanggezhong committed
27
提供基于fastpt不转码编译:
zhanggezhong's avatar
zhanggezhong committed
28

zhanggezhong's avatar
zhanggezhong committed
29
1. 基于光源pytorch基础镜像环境:镜像下载地址:[光合开发者社区](https://sourcefind.cn/#/image/dcu/pytorch),根据pytorch、python、dtk及系统下载对应的镜像版本。
zhanggezhong's avatar
zhanggezhong committed
30

zhanggezhong's avatar
zhanggezhong committed
31
2. 基于现有python环境:安装pytorch,fastpt whl包下载目录:[光合开发者社区](https://sourcefind.cn/#/image/dcu/pytorch),根据python、dtk版本,下载对应pytorch的whl包。安装命令如下:
zhanggezhong's avatar
zhanggezhong committed
32
```shell
zhanggezhong's avatar
zhanggezhong committed
33
34
35
pip install torch* (下载torch的whl包)
pip install fastpt* --no-deps (下载fastpt的whl包, 安装顺序,先安装torch,后安装fastpt)
pip install setuptools==59.5.0 wheel
zhanggezhong's avatar
zhanggezhong committed
36
37
```

zhanggezhong's avatar
zhanggezhong committed
38
39
#### 源码编译安装
- 代码下载
zhanggezhong's avatar
zhanggezhong committed
40
```shell
zhanggezhong's avatar
zhanggezhong committed
41
42
43
44
45
46
47
48
git clone https://developer.sourcefind.cn/codes/OpenDAS/maskrcnn.git # 根据编译需要切换分支
```
- 提供2种源码编译方式(进入maskrcnn目录):
```
1. 设置不转码编译环境变量
source /usr/local/bin/fastpt -C

2. 编译whl包并安装
zhanggezhong's avatar
zhanggezhong committed
49
cd NVIDIA/benchmarks/maskrcnn/implementations/pytorch/maskrcnn
zhanggezhong's avatar
zhanggezhong committed
50
51
52
53
54
python3 setup.py -v bdist_wheel
pip install dist/maskrcnn*

3. 源码编译安装
python3 setup.py install
55
```
zhanggezhong's avatar
zhanggezhong committed
56
57
58
59
60
61
#### 注意事项
+ 若使用pip install下载安装过慢,可添加pypi清华源:-i https://pypi.tuna.tsinghua.edu.cn/simple/
+ ROCM_PATH为dtk的路径,默认为/opt/dtk
+ 在pytorch2.5.1环境下编译需要支持c++17语法,打开setup.py文件,把文件中的 -std=c++14 修改为 -std=c++17

## 验证
62
63
64
65
66
67
68
69
70
71
72
73
74
使用时执行
```shell
source /usr/local/bin/fastpt -e
```

```shell
python3
Python 3.10.12 (main, Feb  4 2025, 14:57:36) [GCC 11.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import maskrcnn_benchmark
>>> maskrcnn_benchmark.__version__
0.1
>>>
zhanggezhong's avatar
zhanggezhong committed
75
```
zhanggezhong's avatar
zhanggezhong committed
76
版本号与官方版本同步,查询该软件的版本号,例如0.1;
zhanggezhong's avatar
zhanggezhong committed
77
78

## Known Issue
zhanggezhong's avatar
zhanggezhong committed
79
-
zhanggezhong's avatar
zhanggezhong committed
80

zhanggezhong's avatar
zhanggezhong committed
81
## 参考资料
zhanggezhong's avatar
zhanggezhong committed
82
- [README_ORIGIN](README_ORIGIN.md)
zhanggezhong's avatar
zhanggezhong committed
83
84
- [README_zh-CN](README_zh-CN.md)
- [https://github.com/mlcommons/training_results_v3.1/tree/main/NVIDIA/benchmarks/maskrcnn/implementations/pytorch/maskrcnn](https://github.com/mlcommons/training_results_v3.1/tree/main/NVIDIA/benchmarks/maskrcnn/implementations/pytorch/maskrcnn)