Install TensorRT

We install TensorRT on RTX2080Ti with CUDA10.0. If you fail to install it, please use pytorch version.

1. install pycuda

export C_INCLUDE_PATH=/usr/local/cuda-10.0/include/:${C_INCLUDE_PATH}
export CPLUS_INCLUDE_PATH=/usr/local/cuda-10.0/include/:${CPLUS_INCLUDE_PATH}
pip install pycuda

2. download tensorrt

• Go to NVIDIA-TENSORRT and then click Download Now.
• Login and download TensorRT7 (please select the version that suits for your platform). We use TensorRT 7.0.0.11 for Ubuntu 18.04 and CUDA 10.0 tar package in our experiment.

3. install

tar -zxf TensorRT-7.0.0.11.Ubuntu-18.04.x86_64-gnu.cuda-10.0.cudnn7.6.tar.gz
vim ~/.bashrc

# Add codes in your file ~/.bashrc
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<your-TensorRT-lib-path>
# for example
# export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/data/zpzhang/TensorRT-7.0.0.11/lib

source ~/.bashrc
conda activate OceanOnlineTRT
cd TensorRT-7.0.0.11/python
# Remember declare your python version=3.7
pip install tensorrt-7.0.0.11-cp37-none-linux_x86_64.whl
cd TensorRT-7.0.0.11/graphsurgeon
pip install graphsurgeon-0.4.1-py2.py3-none-any.whl

4. Verify the installation

python
import tensorrt

5. Install Torch2trt

conda activate OceanOnlineTRT
git clone https://github.com/NVIDIA-AI-IOT/torch2trt
cd torch2trt
python setup.py install

Verify the installation

python
import torch2trt

Note

• If you met the error ImportError: libcudart.so.10.0: cannot open shared object file: No such file or directory, please run sudo cp /usr/local/cuda-10.0/lib64/libcudart.so.10.0 /usr/local/lib/libcudart.so.10.0 && sudo ldconfig.

• If you met the error PermissionError: [Errno 13] Permission denied: '/tmp/torch_extensions/_prroi_pooling/lock', please remove /tmp/torch_extensions/ _prroi_pooling and rerun the tracker. If other user in your machine have compiled prroi pooling before, this may happens. Besides, if you have compiled pproi_pooling before, please remove /tmp/torch_extensions/. Otherwise, you may fail to compile in the new conda environment.