It takes me a lot of time to get TensorRT working with Ubuntu 20.04 on my laptop.
There are some issues makes it even harder @@
1.With the default nVidia driver from Ubuntu 20.04, the laptop failed to resume after suspend (hibernate).
The solution is to reinstall older version. This lead cuda version limited to 10.2
sudo apt purge nvidia-*
sudo apt autoremove
sudo apt install nvidia-driver-450-server
sudo apt install nvidia-driver-450-server
2.Ubuntu 20.04 default is python3.8 and TensorRT works with python3.6
3.TensorRT doesn't support Ubuntu 20.04 with cuda 10.2
The solution is to use python virtualenv to install TensorRT
sudo apt install python3.6-venv
mkdir venv/
cd venv
### Create virtual environment in path venv/tensorrt
python3.6 -m venv tensorrt
source tensorrt/bin/activate
pip install --upgrade pip
python3 -m pip install numpy onnx
### Download & extract TensorRT-7.2.3.4.Ubuntu-18.04.x86_64-gnu.cuda-10.2.cudnn8.1.tar.gz
cd Downloads/
sudo cp -a TensorRT-7.2.3.4 /usr/local/
export LD_LIBRARY_PATH=/usr/local/TensorRT-7.2.3.4/lib
cd TensorRT-7.2.3.4/python/
python3 -m pip install tensorrt-7.2.3.4-cp36-none-linux_x86_64.whl
cd ../uff
python3 -m pip install uff-0.6.9-py2.py3-none-any.whl uff-0.6.9-py2.py3-none-any.whl
which convert-to-uff
cd ../graphsurgeon/
python3 -m pip install graphsurgeon-0.4.5-py2.py3-none-any.whl
cd ../onnx_graphsurgeon/
python3 -m pip install onnx_graphsurgeon-0.2.6-py2.py3-none-any.whl
cd ../..
### Download libcudnn8_8.2.1.32-1+cuda10.2_amd64.deb & libcudnn8-dev_8.2.1.32-1+cuda10.2_amd64.deb
sudo dpkg -i ./libcudnn8_8.2.1.32-1+cuda10.2_amd64.deb
sudo dpkg -i ./libcudnn8-dev_8.2.1.32-1+cuda10.2_amd64.deb
pip3 install torch
pip3 install torchvision
pip3 install matplotlib
pip3 install --global-option=build_ext --global-option="-I/usr/local/cuda-10.2/targets/x86_64-linux/include/" --global-option="-L/usr/local/cuda-10.2/targets/x86_64-linux/lib/" pycuda
pip3 install opencv-python
pip3 install albumentations==0.5.2
4.Using the tensorrt venv
cd venv
source tensorrt/bin/activate
沒有留言:
張貼留言