detectron2 windows 安装

https://github.com/facebookresearch/detectron2官方目前暂时没有提供官方的windows安装教程，这里记录使用中碰到的坑

1. 主要参考

https://github.com/conansherry/detectron2

2. 安装依赖

1. 安装已有的包

   conda create -n detectron2 python=3.8
   
   conda activate detectron2
   
   conda install pytorch torchvision cudatoolkit=10.2 -c pytorch
   
   pip install cython pycocotools-windows fvcore pydot future tensorboard tqdm mock matplotlib cloudpickle tabulate yacs Pillow termcolor opencv-python
   

2. 安装vs2019或者2017，然后执行如下
   2019版本执行
   set DISTUTILS_USE_SDK=1 set MAX_JOBS=5 "H:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Auxiliary\Build\vcvars64.bat"
   2017版本执行
   set DISTUTILS_USE_SDK=1 set MAX_JOBS=5 "C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Auxiliary\Build\vcvars64.bat"

3. 修改几处代码

   file1: 
      E:\ProgramData\Miniconda3\envs\detectron2\lib\site-packages\torch\include\torch\csrc\jit\argument_spec.h
     example:
     {C:\Miniconda3\envs\py36}\Lib\site-packages\torch\include\torch\csrc\jit\argument_spec.h(190)
       static constexpr size_t DEPTH_LIMIT = 128;
         change to -->
       static const size_t DEPTH_LIMIT = 128;
   
   file2: 
      E:\ProgramData\Miniconda3\envs\detectron2\lib\site-packages\torch\include\pybind11\cast.h
     example:
     {C:\Miniconda3\envs\py36}\Lib\site-packages\torch\include\pybind11\cast.h(1449)
       explicit operator type&() { return *(this->value); }
         change to -->
       explicit operator type&() { return *((type*)this->value); }
   
   file3:
    {detectron2 src path}\detectron2\layers\csrc\cocoeval\cocoeval.cpp
   注释掉下面几行 localtime_r只有linux有
   //localtime_r(&rawtime, &local_time);
   //strftime(
   //    buffer.data(), 200, "%Y-%m-%d %H:%num_max_detections:%S", &local_time);
   
   

4. 下载源码

   git clone --depth=1 https://hub.fastgit.org/facebookresearch/detectron2.git
   cd detectron2
   python setup.py build develop
   

5. [选择执行，如果后面运行没有win32错误]重新安装pywin32

   pip uninstall pywin32
   pip install pywin32==225
   

6. 安装成功