Tensorrt环境搭建

Tensorrt环境搭建

环境安装

python 3.6
TensorRT 7.0.0.1
https://developer.nvidia.com/nvidia-tensorrt-7x-download

cd /path/to/TensorRT7.0.0.1
pip install tensorrt-7.0.0.11-cp36-none-linux_x86_64.whl

pytorch 1.5
https://pytorch.org/get-started/locally/
根据机器需求,下载匹配的pytorch
cudnn 7.6
https://developer.nvidia.com/rdp/cudnn-archive
下载后,将其解压,放置cuda安装的位置即可
cuda 10.2
https://developer.nvidia.com/cuda-toolkit-archive

wget http://developer.download.nvidia.com/compute/cuda/10.2/Prod/local_installers/cuda_10.2.89_440.33.01_linux.run
sudo sh cuda_10.2.89_440.33.01_linux.run

编辑~/.bashrc

export PATH=$PATH:/usr/local/cuda-10.2/bin
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-10.2.0/lib64

cuda安装注意事项 一定不要安装默认的驱动,而且options里面也不需要编译opengl的库

nvidia驱动安装
搜索适合的驱动,然后下载

https://www.nvidia.cn/Download/index.aspx?lang=cn

pycuda

pip install  pycuda -i http://mirrors.aliyun.com/pypi/simple/  --trusted-host mirrors.aliyun.com

options

pycharm
https://www.jetbrains.com/pycharm/download/#section=linux
opencv

pip install  opencv-python==3.4.2.17 -i http://mrrors.aliyun.com/pypi/simple/  --trusted-host mirrors.aliyun.com

nvidia-docker安装
https://www.jianshu.com/p/ef8b0e6c6f5c

trubble-shooting

1. src/cpp/cuda.hpp:14:18: fatal error: cuda.h: No such file or directory
   可能是没有安装cuda或者是安装cuda后include没有加入PATH
2. pycharm服务器运行没有图形界面,加入-X

ssh -X [email protected]

3. pip install老是安装到其他路径了,可以指定安装的位置

sudo ~/anaconda3/bin/python setup.py install

4.查看显卡使用情况

watch nvidia-smi

其他

CUDA GPU算力对照表:
https://developer.nvidia.com/cuda-gpus#collapseOne

Tensorrt C++源码库地址
https://github.com/NVIDIA/TensorRT

sudo cmake .. -DTRT_LIB_DIR=$TRT_RELEASE/lib -DTRT_BIN_DIR=`pwd`/out -DCMAKE_CUDA_COMPILER:PATH=/usr/local/cuda/bin/nvcc