目录
一、环境版本
1.1 概述
系统环境：Ubuntu18.04 python=3.8.5 torch=1.10.0 TensorRT=7.2.3.4 opencv=4.5.1
cuda=11.2 cudnn=8.2.2 Driver Version: 460.32.03 显卡：RTX 3090
1.2 Ubuntu版本
在终端运行： lsb_release -a
1.3 python、pytorch与tensorrt版本
在终端运行： lsb_release -a
1.4 cudnn版本
在终端运行： dpkg -L libcudnn8
1.5 驱动版本
在终端运行： watch -n 1 nvidia-smi
1.6 opencv版本
在终端运行： pkg-config --modversion opencv
二、TensorRT与opencv环境安装
2.1 opencv安装
- 可参考网：Home - OpenCV
- ippicv包：github.com/opencv/open…
- 编译命令：cmake -D CMAKE_INSTALL_PREFIX=/usr/local -D CMAKE_BUILD_TYPE=Release -D BUILD_TIFF=ON
注：在conda环境中，由于安装pytorch，会自动安装“libtiff”，卸载后导致pytorch不能用，然后导致opencv在编译过程中会报错，具体如下：
/usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFReadRGBAStrip@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFReadDirectory@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFWriteEncodedStrip@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFIsTiled@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFWriteScanline@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFGetField@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFScanlineSize@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFWriteDirectory@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFReadEncodedTile@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFReadRGBATile@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFClose@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFClientOpen@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFRGBAImageOK@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFOpen@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFReadEncodedStrip@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFSetField@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFSetWarningHandler@LIBTIFF_4.0' /usr/local/lib/libopencv_imgcodecs.so.4.5.1: undefined reference to `TIFFSetErrorHandler@LIBTIFF_4.0
解决方法：
- opencv源码安装编译时候，保证conda环境没有libtiff包名
- 在cmake 编译过程中开启 BUILD_TIFF
参考链接：
- LIBTIFF_4.0 link errors - OpenCV Q&A Forum
- stackoverflow.com/questions/2…
- OpenCV LIBTIFF_4.0 link errors - 普兒 - 博客园
- Linux（Ubuntu 18.04）中安装OpenCV + OpenCV_Contrib - 花落未殇 - 博客园
- blog.51cto.com/u_14587161/…
2.2 TensorRT安装
- 可参考网：developer.nvidia.com/zh-cn/tenso…，有tar包和deb两种安装方，本文选择tar包进行解压安装。
- 根据tar包解压安装：tar -xvf TensorRT-7.2.3.4.Ubuntu-18.04.x86_64-gnu.cuda-11.1.cudnn8.1.tar.gz
- 安装命令：分别进入python、uff和graphsurgeon根据python环境，进行pip安装相应的包
- vi ~/.bashrc 添加 export LD_LIBRARY_PATH=/usr/local/TensorRT-7.2.3.4/lib:/usr/local/cuda/bin:INCLUDE_PATH
- source ~/.bashrc
三、yolov5转tensorrt
3.1 下载yolov5项目
3.2 下载tensorrtx项目
3.3 开始准备编译
在tensorrx项目中，进入yolov5目录，修改Cmakelists.txt:
修改自己tcuda和tensorrt路径：
# cuda
include_directories(/usr/local/cuda/include)
link_directories(/usr/local/cuda/lib64)
# tensorrt
include_directories(/usr/local/TensorRT-7.2.3.4/include)
link_directories(/usr/local/TensorRT-7.2.3.4/lib)
3.4 获取wts模型
在终端运行： ****python gen_wts.py
3.5 编译命令
在终端运行：mkdir build && cd build:
在终端运行：cmake ..
在终端运行：make -j128
在终端运行：./yolov5 -s // serialize model to plan file i.e. 'yolov5s.engine' 序列化模型
3.6 测试
在终端运行：./yolov5 -d ../data/images // deserialize plan file and run inference, the images in samples will be processed. 测试用例
或者在终端运行：CUDA_VISIBLE_DEVICES=6 python yolov5_trt.py
