MEGA配置和在单GPU训练

环境安装：

可以参考install.md文档，和 maskrcnn_benchmark 的 install.md 文档

 

在rfbnet环境下测试是否可以运行MEAG程序

 

         conda install ipython pip

        

         pip install ninja yacs cython matplotlib tqdm opencv-python scipy

        

         export INSTALL_DIR=$PWD

 

         # install pycocotools

         cd $INSTALL_DIR

         git clone https://github.com/cocodataset/cocoapi.git

         发现未安装git，手动下载

         cd cocoapi/PythonAPI

         python setup.py build_ext install

        

         # install apex

         cd $INSTALL_DIR

         git clone https://github.com/NVIDIA/apex.git

         未安装git，手动下载

         cd apex

         python setup.py install --cuda_ext --cpp_ext

         报错：not find files "usr\local\cuda\bin\nvcc"

        

         查看cuda位置

        

         export CUDA_HOME = usr\local\cuda\cuda-10.0

         cd apex

         python setup.py install --cuda_ext --cpp_ext

         报错：Cuda版本和pytorch二进制文件的cuda不匹配

        

         查看CUDA版本10.0

         torch.version.cuda:9.0.176

         需让他们版本一致

        

创建新的虚拟环境，重新测试

 

         conda create --name MEGA -y python=3.7

         source activate MEGA

 

         # this installs the right pip and dependencies for the fresh python

         conda install ipython pip

 

         # mega and coco api dependencies

         pip install ninja yacs cython matplotlib tqdm opencv-python scipy

 

         # follow PyTorch installation in https://pytorch.org/get-started/locally/

         # we give the instructions for CUDA