AI开箱 C++调用Detectron2的Mask R-CNN 欢迎订阅我的频道

• bilibili频道
• youtube频道

视频

• 为保障项目复现，本视频在虚拟机下录制，系统: ubuntu-18.04.5-desktop-amd64.iso
• 虚拟机磁盘空间，建议分配40G

bilibili
[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-16nnXWoo-1606283199472)(https://github.com/dyh/unbox_use_detectron2_model_in_cpp_of_caffe2/blob/main/cover.png?raw=true)]
系统需求

• ubuntu 18.04
• python 3.6
• cuda 10.1
• cudnn 8.0.5

第一部分：转换为 caffe2 模型 python 程序环境配置

1. 下载代码
   安装 git
   

   sudo apt install git

   
   下载代码
   

   git clone https://github.com/dyh/unbox_use_detectron2_model_in_cpp_of_caffe2.git unbox_cpp_caffe2

   
   
2. 进入目录
   

   cd unbox_cpp_caffe2/python_project/

   
   
3. 创建 python 虚拟环境
   

   sudo apt install python3-venv python3 -m venv venv

   
   
4. 激活虚拟环境
   

   source venv/bin/activate

   
   
5. 升级pip
   

   python -m pip install --upgrade pip

   
   
6. 安装软件包
   
   1. 安装gcc
      

      sudo apt install gcc

      
      
   2. 安装CUDA和CUDNN
      

      下载 http://developer.download.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.243_418.87.00_linux.run sudo bash cuda_10.1.243_418.87.00_linux.run下载 https://developer.nvidia.com/compute/machine-learning/cudnn/secure/8.0.5/10.1_20201106/cudnn-10.1-linux-x64-v8.0.5.39.tgz 解压，将include目录和lib64目录下的文件拷贝至 /usr/local/cuda 对应目录

      
      
   3. 安装 torch==1.6.0+cu101
      

      pip install torch==1.6.0+cu101 torchvision==0.7.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html

      
      
   4. 安装 detectron2==0.3+cu101
      

      sudo apt install python3.6-dev python -m pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/torch1.6/index.html

      
      
   5. 安装其他包
      

      sudo apt-get install graphviz pip install opencv-python==4.4.0.46 pip install onnx==1.8.0 pip install protobuf==3.14.0

      
      
7. 下载训练好的 Mask R-CNN weights 文件
   

   model_0124999.pth，334.8 MB，这是我在上一期视频中，使用Mask R-CNN检测隧道裂缝的权重文件，包含2个分类：裂缝fissure和渗水water
   

   • 下载链接1: https://pan.baidu.com/s/1BqUTgciTeDxng21dYcMlag 提取码: puaq
   • 下载链接2: https://drive.google.com/file/d/1SLs-dCHibUMJY0dgcbkAb82h-Yxm6gAs/view?usp=sharing
   • 形成 ./python_project/weights/model_0124999.pth 的目录结构
8. 关于训练图片和标注文件
   

   python_project/images/train 目录中已经包含训练图片和标注文件，关于如何制作这个JSON标注文件，请参考这个视频：
   

   • bilibili
   • youtube

运行 python 程序转换模型

python caffe2_converter.py

程序运行成功后，将在 python_project/output 目录生成 model.pb 和 model_init.pb 文件
第二部分：使用 C++ 调用模型 C++ 程序环境配置

1. 进入目录
   

   cd ../cpp_project/

   
   
2. 安装依赖项
   

   sudo apt install libgflags-dev libgoogle-glog-dev libopencv-dev pip install mkl-include

   
   
3. 安装 protobuf
   

   下载 https://github.com/protocolbuffers/protobuf/releases/download/v3.11.4/protobuf-cpp-3.11.4.tar.gz tar xf protobuf-cpp-3.11.4.tar.gz cd protobuf-3.11.4 export CXXFLAGS=-D_GLIBCXX_USE_CXX11_ABI=$(python3 -c 'import torch; print(int(torch.compiled_with_cxx11_abi()))') ./configure --prefix=$HOME/.local && make && make install

   
   
4. 配置 CMakeLists.txt
   

   回到 cpp_project 目录，修改 CMakeLists.txt 文件中的内容
   

   1. 配置 pytorch 路径
      

      修改路径 set(CMAKE_PREFIX_PATH $ENV{HOME}/workspace/unbox/unbox_cpp_caffe2/python_project/venv/lib64/python3.6/site-packages/torch) 指向 python 虚拟环境中的 pytorch 安装目录

      
      
   2. 配置 protobuf 路径
      

      # point to the include folder of protobuf include_directories($ENV{HOME}/.local/include) # point to the lib folder of protobuf link_directories($ENV{HOME}/.local/lib)

      
      

运行 C++ 程序检测图片

1. 编译
   

   回到 cpp_project 目录
   

   安装cmake
   

   sudo apt install cmake mkdir build && cd build cmake .. && make

   
   
2. 运行
   

   ./caffe2_mask_rcnn

   
   

官方参考

• https://github.com/facebookresearch/detectron2/tree/master/tools/deploy
• https://detectron2.readthedocs.io/tutorials/deployment.html

Unbox AI, use Mask R-CNN of detectron2 in C++ welcome to subscribe my channel

• youtube channel
• bilibili channel

Video

• this video is recorded under the virtual machine (ubuntu-18.04.5-desktop-amd64.iso) to ensure the recurrence of the project
• it is recommended to allocate 40G disk space of virtual machine

youtube
[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-548FEoU5-1606283199474)(https://github.com/dyh/unbox_use_detectron2_model_in_cpp_of_caffe2/blob/main/cover.png?raw=true)]
System Requirements

• ubuntu 18.04
• python 3.6
• cuda 10.1
• cudnn 8.0.5

Chapter 1: Convert to Caffe2 Model Python Environment configuration

1. download source code
   install git
   

   sudo apt install git

   
   clone source code to unbox_cpp_caffe2 folder
   

   git clone https://github.com/dyh/unbox_use_detectron2_model_in_cpp_of_caffe2.git unbox_cpp_caffe2

   
   
2. enter the python project folder
   

   cd unbox_cpp_caffe2/python_project/

   
   
3. create the virtual environment
   

   sudo apt install python3-venv python3 -m venv venv

   
   
4. activate virtual environment
   

   source venv/bin/activate

   
   
5. upgrade pip
   

   python -m pip install --upgrade pip

   
   
6. install software packages
   
   1. install gcc
      

      sudo apt install gcc

      
      
   2. install CUDA and CUDNN
      

      download file: http://developer.download.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.243_418.87.00_linux.run sudo bash cuda_10.1.243_418.87.00_linux.rundownload file: https://developer.nvidia.com/compute/machine-learning/cudnn/secure/8.0.5/10.1_20201106/cudnn-10.1-linux-x64-v8.0.5.39.tgz unzip files, copy the files of include and lib64 to /usr/local/cuda/include and /usr/local/cuda/lib64 folder

      
      
   3. install torch==1.6.0+cu101
      

      pip install torch==1.6.0+cu101 torchvision==0.7.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html

      
      
   4. install detectron2==0.3+cu101
      

      sudo apt install python3.6-dev python -m pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/torch1.6/index.html

      
      
   5. install other software packages
      

      sudo apt-get install graphviz pip install opencv-python==4.4.0.46 pip install onnx==1.8.0 pip install protobuf==3.14.0

      
      
7. download pre-trained weights file of Mask R-CNN
   

   model_0124999.pth, 334.8MB, in my last video, this is a weights file that uses Mask R-CNN to detect tunnel fissure, including 2 categories: fissure and water
   

   • download link 1: https://drive.google.com/file/d/1SLs-dCHibUMJY0dgcbkAb82h-Yxm6gAs/view?usp=sharing
   • download link 2: https://pan.baidu.com/s/1BqUTgciTeDxng21dYcMlag password: puaq
   • make the directory structure as ./python_project/weights/model_0124999.pth
8. About sample images and annotation files
   

   The python_project/images/train directory already contains sample images and annotation files. You can refer to this video on how to make this annotation JSON file:
   

   • youtube
   • bilibili

Run Python program to convert the model file & weights file

python caffe2_converter.py

when the program runs successfully, model.pb and model_init.pb files are generated in the python_project/output directory
Chapter 2: Use C++ to load weights file C++ Environment Configuration

1. enter the C++ project folder
   

   cd ../cpp_project/

   
   
2. install dependency packages
   

   sudo apt install libgflags-dev libgoogle-glog-dev libopencv-dev pip install mkl-include

   
   
3. install protobuf
   

   download file: https://github.com/protocolbuffers/protobuf/releases/download/v3.11.4/protobuf-cpp-3.11.4.tar.gz tar xf protobuf-cpp-3.11.4.tar.gz cd protobuf-3.11.4 export CXXFLAGS=-D_GLIBCXX_USE_CXX11_ABI=$(python3 -c 'import torch; print(int(torch.compiled_with_cxx11_abi()))') ./configure --prefix=$HOME/.local && make && make install

   
   
4. configure the CMakeLists.txt
   

   return to the cpp_project directory and modify the contents of the CMakeLists.txt file
   

   1. configure pytorch path
      

      change path: set(CMAKE_PREFIX_PATH $ENV{HOME}/workspace/unbox/unbox_cpp_caffe2/python_project/venv/lib64/python3.6/site-packages/torch) Point to the pytorch installation directory in the python virtual environment

      
      
   2. configure protobuf path
      

      # point to the include folder of protobuf include_directories($ENV{HOME}/.local/include) # point to the lib folder of protobuf link_directories($ENV{HOME}/.local/lib)

      
      

Run C++ Program to Detect Images

1. compile
   

   return to the cpp_project directory
   

   install cmake
   

   sudo apt install cmake mkdir build && cd build cmake .. && make

   
   
2. 【Unbox|C++ 调用 Mask R-CNN Detectron2】run
   

   ./caffe2_mask_rcnn

   
   

Official Reference

• https://github.com/facebookresearch/detectron2/tree/master/tools/deploy
• https://detectron2.readthedocs.io/tutorials/deployment.html

推荐阅读

• 

  北控水务为什么不涨 北控水务到底好不好
• 

  牵组词语 汉字牵组词语
• 

  月季花扦插的最佳时间 月季花扦插的最佳时间现在可以插吗?
• 

  2023重庆医科大学研究生成绩复查申请时间+方式
• 

  n95口罩能防尘吗 n95口罩可以用几次
• 

  摩托罗拉里程碑2刷机
• 

  如果你爱的人总是以怨报德,你会怎么样？
• 

  宝马525多少钱落地 宝马525价格多少钱
• 

  电商铺货是什么意思 电商货架什么意思，货架市场
• 

  胡萝卜怎样吃最有营养啊视频 胡萝卜怎样吃最有营养啊
• 

  巴戟天药材多少钱一斤，巴戟天药店多少钱一斤
• 

  吕锅上的污渍怎么去除啊
• 

  非主流图片唯美霸气 非主流相框素材
• 

  白金膏的功效与作用 皮白金膏的功效与作用?
• 

  虚拟机怎么开网页游戏模式，虚拟机怎样打开网页
• 

  闻鸡起舞的主人公是谁 囊萤映雪的主人公是谁
• 

  口红吃饭的时候怎么办
• 

  宝宝长期咳嗽易得肺炎
• 

  包贝尔大婚柳岩遭闹伴娘 柳岩伴娘事件是什么?
• 

  未婚生子罚款多少2022年 未婚生子罚款多少2022？

• C++从入门到精通|C++（ 指针微探（指针与数组））
• 使用30分钟快速入门机器学习时间
• 10个基本的机器学习面试问题分享
• Facebook使用机器学习的5种令人振奋的方式
• python|从零开始用 Python 构建一个简单的神经网络
• python机器学习|从零到一实现神经网络（python）（一）
• 如何使用Jupyter Notebook（-终极指南）
• 人工智能|浙大吴飞「舌战」阿里贾扬清（AI内卷与年薪百万，哪个才是真实（））
• Flink|独家专访阿里集团副总裁贾扬清（我为什么选择加入阿里巴巴（））