1. First, first list the required installation environment

Each library version number

cuda10.2.89+cudnn7.6.5

pytorch 1.8.1

conda install pytorch==1.8.1 cudatoolkit==10.2.89 -c pytorch

Slowly installation can refer toConda Install PyTorch installation solution -Zhihu

conda install cudnn==7.6.5

pip install torchvision==0.9.1

pip install onnx==1.8.0

pip install onnxruntime==1.6.0

Download tensorrt7.2.1

https://developer.nvidia.com/nvidia-tensorrt-7x-download

Download 7.1FOR Linux and CUDA10.2

cd /home/wx/Downloads/TensorRT-7.2.1.6.Ubuntu-18.04.x86_64-gnu.cuda-10.2.cudnn8.0/TensorRT-7.2.1.6/python

pip install tensorrt-7.2.1.6-cp37-none-linux_x86_64.whl

Download onnx-Tensorrt7.2

refer to [Tensorrt series] 2.onnx -Tensorrt installation tutorial -Zhihu

git clone --recursive -b 7.2.1 https://github.com/onnx/onnx-tensorrt.git onnx-tensorrt7.2

Configure the Tensorrt path and CUDA path. CUDA's path in the environment created by Conda is not a local CUDA path

cmake .. -DTENSORRT_ROOT=//home/wx/Downloads/TensorRT-7.2.1.6.Ubuntu-18.04.x86_64-gnu.cuda-10.2.cudnn8.0/TensorRT-7.2.1.6 -DCUDA_TOOLKIT_ROOT_DIR=/home/wx/anaconda3/envs/solo-3 -DONNX_BUILD_TESTS=ON

When compiling and installation, you need propobuf> = 3.19.0

pip install protobuf==3.19.0

Finally compile Solov2.Tensorrt

Error 1

Error cpython-37m-x86_64-linux-Gnu.so: Undefined Symbol: ThPVARIableClasss

Solution: 1. Check the version matching 2. It is compiled before, delete Build re -compilation

train error cpython-37m-x86_64-linux-gnu.so: undefined symbol: THPVariableClass

Error 2

nvcc fatal   : Unknown option '-generate-dependencies-with-compile'

Find the last line of setup.py,

Original cmdclass = {'Build_ext': BuildextEntension},
        zip_safe=False)

Modify to cmdclass = {'Build_ext': Buildextent.With_options (use_ninja = false)},
        zip_safe=False)

Error 3. Error: ‘AT_CHECK’ WAS Not Declared in this scope

PyTorch version <1.5 will appear. Modify AT_CHECK as torch_check

Error 4 NVCC FATAL: Redefinition of Argument 'STD'

Modify cmakelists.txt, comment on the list (APPEND CUDA_NVCC_FLAGS "-Std = C ++ 11")

Most of the errors are caused by the version that is not matched. I tried several versions of PyTorch1.1 / 1.4 / 1.11

CUDA 10.0 10.1 10.2 11.3 Trial several times before running on this version. After installing the environment, you can run deploy/quick_onnx_trt_test.py to see if the environment is right.