目录
前言
1--安装Anaconda
2--创建conda环境及安装pytorch
3--配置环境依赖
4--安装mmdeploy
5--实例测试
6--参考
前言
建议结合官方文档阅读。
1--安装Anaconda
①下载Anaconda
cd ~/mmdeploy_test # 进入一个新的文件夹 # 下载anaconda包 wget "https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge-pypy3-Linux-aarch64.sh" bash Miniforge-pypy3-Linux-aarch64.sh # 执行安装 source ~/.bashrc # 更新Path
②设置国内源
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/ conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/ conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/ conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/peterjc123/ conda config --add channels http://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/ conda config --add channels http://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/ conda config --add channels http://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/ conda config --add channels http://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/peterjc123/ conda config --set show_channel_urls yes
③测试
conda config --show channels # 查看源路径 conda --version # 查看版本
2--创建conda环境及安装pytorch
①创建mmdeploy环境，python版本为3.6
# 创建mmdeploy环境 conda create -n mmdeploy python=3.6 -y conda activate mmdeploy
python版本建议安装3.6和3.8，以与nvidia jetson提供的pytorch安装包适配。
②下载pytorch安装包
国内无pvn的情况下，建议从nvidia jetson提供的pytorch官网下载pytorch安装包。
# 在windows或服务器上下载pytorch安装包 # 通过sftp传到nvidia jetson中 # 以代替命令：wget https://nvidia.box.com/shared/static/fjtbno0vpo676a25cgvuqc1wty0fkkg6.whl -O torch-1.10.0-cp36-cp36m-linux_aarch64.whl
③安装pytorch
# 安装依赖库 sudo apt-get install libopenblas-base libopenmpi-dev libomp-dev pip3 install Cython numpy # 运行pytorch安装包 pip3 install torch-1.10.0-cp36-cp36m-linux_aarch64.whl
④添加路径
# 添加路径 vim ~/.bashrc export OPENBLAS_CORETYPE=ARMV8 # 增加这行内容 :wq # 保存退出 source ~/.bashrc
⑤安装torchvision（这个过程可能会卡顿或者持续时间较长）
# 安装torchvision sudo apt-get install libjpeg-dev zlib1g-dev libpython3-dev libavcodec-dev libavformat-dev libswscale-dev -y git clone https://github.com/pytorch/vision torchvision cd torchvision git checkout tags/v0.11.1 -b v0.11.1 export BUILD_VERSION=0.11.1 pip install -e .
3--配置环境依赖
①配置cmake
## 更新cmake
# 卸载已安装的版本
sudo apt-get purge cmake
sudo snap remove cmake
# 安装新版本
export CMAKE_VER=3.23.1
export ARCH=aarch64
wget https://github.com/Kitware/CMake/releases/download/v${CMAKE_VER}/cmake-${CMAKE_VER}-linux-${ARCH}.sh
chmod +x cmake-${CMAKE_VER}-linux-${ARCH}.sh
sudo ./cmake-${CMAKE_VER}-linux-${ARCH}.sh --prefix=/usr --skip-license
cmake --version
②配置Tensorrt
配置Tensorrt cp -r /usr/lib/python3.6/dist-packages/tensorrt* /home/nvidia/miniforge-pypy3/envs/mmdeploy/lib/python3.6/site-packages/ conda deactivate conda activate mmdeploy python -c "import tensorrt; print(tensorrt.__version__)" # 测试TensorRT版本 # 设置路径 利用vim ~/.bashrc进行 export TENSORRT_DIR=/usr/include/aarch64-linux-gnu export PATH=$PATH:/usr/local/cuda/bin export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64
③安装mmcv-full （这个过程编译可能会卡顿或者持续时间较长）
# 安装mmcv-full sudo apt-get install -y libssl-dev git clone https://github.com/open-mmlab/mmcv.git cd mmcv git checkout -b v1.4.0 MMCV_WITH_OPS=1 pip install -e .
④安装onnx
pip install onnx
⑤安装h5py
sudo apt-get install -y pkg-config libhdf5-100 libhdf5-dev # 安装依赖 pip install versioned-hdf5 # 安装
⑥安装spdlog
sudo apt-get install -y libspdlog-dev
⑦安装ppl.cv
git clone https://github.com/openppl-public/ppl.cv.git # 下载 cd ppl.cv export PPLCV_DIR=$(pwd) echo -e 'n# set environment variable for ppl.cv' >> ~/.bashrc # 设置path echo "export PPLCV_DIR=$(pwd)" >> ~/.bashrc git checkout tags/v0.6.2 -b v0.6.2 ./build.sh cuda
⑧安装mmdetecttion
4--安装mmdeploy
①下载mmdeploy
git clone --recursive https://github.com/open-mmlab/mmdeploy.git cd mmdeploy export MMDEPLOY_DIR=$(pwd)
②编译TensorRT自定义算子
# 编译tensorRT自定义算子 mkdir -p build && cd build cmake .. -DMMDEPLOY_TARGET_BACKENDS="trt" make -j$(nproc)
③安装mmdeploy依赖库
# 安装MMDeploy依赖库
cd ${MMDEPLOY_DIR}
pip install -v -e .
④编译C++ SDK
mkdir -p build && cd build
cmake ..
-DMMDEPLOY_BUILD_SDK=ON
-DMMDEPLOY_BUILD_SDK_PYTHON_API=ON
-DMMDEPLOY_TARGET_DEVICES="cuda;cpu"
-DMMDEPLOY_TARGET_BACKENDS="trt"
-DMMDEPLOY_CODEBASES=all
-Dpplcv_DIR=${PPLCV_DIR}/cuda-build/install/lib/cmake/ppl
make -j$(nproc) && make install
⑤编译demo sdk（六个.cpp文件）
# 编译demo sdk
cd ${MMDEPLOY_DIR}/build/install/example
mkdir -p build && cd build
cmake .. -DMMDeploy_DIR=${MMDEPLOY_DIR}/build/install/lib/cmake/MMDeploy
make -j$(nproc)
5--实例测试
①mask_RCNN模型，python API模型转换
## mask_RCNN 实例 # 调用pythonAPI 转换模型： pytorch→onnx→engine python tools/deploy.py configs/mmdet/instance-seg/instance-seg_tensorrt_dynamic-320x320-1344x1344.py /home/nvidia/mmdeploy_test/mmdetection/configs/mask_rcnn/mask_rcnn_x101_64x4d_fpn_mstrain-poly_3x_coco.py # 这个要在mmdetection源码中引用 checkpoints/mask_rcnn_x101_64x4d_fpn_mstrain-poly_3x_coco_20210526_120447-c376f129.pth maskrcnn_file/demo.jpg --work-dir work_dirs/maskrcnn/ --device cuda --show --dump-info
引用“/home/nvidia/mmdeploy_test/mmdetection/configs/mask_rcnn/mask_rcnn_x101_64x4d_fpn_mstrain-poly_3x_coco.py”文件时，需要在mmdetection源码中引用，所以要先git clone 下载mmdetection
②C++推理实例
# 测试C++实例 # 运行实例分割程序（参数1：gpu加速； 参数2：推理模型的地址； 参数3：推理测试图片的地址） ./object_detection cuda /home/nvidia/mmdeploy_test/mmdeploy/work_dirs/maskrcnn/ /home/nvidia/mmdeploy_test/mmdeploy/maskrcnn_file/demo.jpg
6--参考
官方安装文档
