锐客网

  1. 首页 > 生活百科 > it技术 > >

Unbox|C++ 调用 Mask R-CNN Detectron2

ai机器学习神经网络深度学习c++pytorchunbox

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


    上一篇：C++从入门到精通|C++（ 指针微探（指针与数组））

    下一篇：Opencv|【Opencv-python】画图操作/画线/画矩形/画圆/画多边形/添加文字