cuda、cudnn、zlib 深度学习GPU必配三件套（Windows）
本文是按照 tensorrtx/yolo11 at master · wang-xinyu/tensorrtx (github.com) 要求的版本进行下载安装，跨大版本不推荐，到处是坑、坑、坑~
默认已经安装了VS2022、英伟达显卡的最新版本驱动。
1、cuda 11.8
下载地址：CUDA Toolkit Archive | NVIDIA 开发者
默认安装在 C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8，注意下图两步，其他默认。
2、cudnn 8.9.7
下载地址：cuDNN Archive | NVIDIA Developer
解压后：
dll 放到 C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin
lib 放到 C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\lib\x64
include 放到 C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\include
之后，可以把cudnn文件删除了。
关于新版本：
新版本开始提供exe安装了，也可以压缩包方式安装。更新太快，建议查看官方文档
8版本 Documentation Archives :: NVIDIA cuDNN Documentation
9版本 Release Notes — NVIDIA cuDNN v9.5.0 documentation
3、zlib
Zlib是cuDNN所需的数据压缩软件库。从cudnn 8.9.4 版本开始，window系统不需要再安装zlib了，已经链接进cudnn的动态库里了。linux还需安装，参见2步骤的官方文档链接。
ZLIB version 1.2.13 is statically linked into the cuDNN Windows dynamic libraries.
具体下载地址查看官方文档 Documentation Archives :: NVIDIA cuDNN Documentation
解压后，
zlibwapi.dll 放到 C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin
zlibwapi.lib 放到 C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\lib\x64
4、tensorrt 8.6
下载地址：TensorRT Download | NVIDIA Developer
GA是稳定版
解压后，直接放到了C盘根目录。
把lib目录添加进系统环境变量，因为dll在此目录里。
5、opencv
下载安装（其实就是解压）到C盘根目录
同上，把bin目录添加进系统环境变量，因为dll在此目录里。
~~~~~~~~~~~~~~~~~ 以上是部署模型要用的，以下是训练模型要用的~~~~~~~~~~~~~~~~~~~
6、minconda3+pyhton3.10+pytorch
训练用python环境训练，推荐装到conda虚拟环境里。
下载地址： Download Anaconda Distribution | Anaconda（打开的网页，往下滑就能看到Miniconda），按照下图安装，其他默认。
miniconda3安装到D盘（怕C盘老是要权限）
打开conda终端，创建环境，指定python3.10版本
conda create -n yolo11 python=3.10
激活yolo11环境
conda activate yolo11
安装pytorch，PyTorch 官网复制粘贴命令行，回车安装即可。
安装tensorrt，其实就是执行下图文件
conda终端里执行：
pip install C:\TensorRT\python\tensorrt-8.6.1-cp310-none-win_amd64.whl
7、训练yolo
conda终端里执行：
pip install ultralytics
pycharm软件，新建python项目，选择yolo11虚拟环境，main.py里输入
from ultralytics import YOLO # Load a model model = YOLO("yolo11n.pt") # Train the model train_results = model.train( data="coco8.yaml", # path to dataset YAML epochs=10, # number of training epochs imgsz=640, # training image size device=0, # device to run on, i.e. device=0 or device=0,1,2,3 or device=cpu batch=1, workers=0 ) # Evaluate model performance on the validation set metrics = model.val() # Perform object detection on an image results = model("https://ultralytics.com/images/bus.jpg") results[0].show() # Export the model to ONNX format # pathOnnx = model.export(format="onnx") # return path to exported model pathEngine = model.export(format="engine", device=0)
报错就多运行几次，因为有一些包会下载失败。
最终生成 yolo11n.engine，说明训练环境OK。
8、部署yolo
下载 tensorrtx/yolo11 at master · wang-xinyu/tensorrtx (github.com)
使用vscode将yolo11编译成vs2022项目，CMakeLists.txt文件内容略微修改如下：
cmake_minimum_required(VERSION 3.10) project(yolov11) add_definitions(-std=c++11) add_definitions(-DAPI_EXPORTS) set(CMAKE_CXX_STANDARD 11) set(CMAKE_BUILD_TYPE Debug) set(CMAKE_CUDA_COMPILER /usr/local/cuda/bin/nvcc) enable_language(CUDA) include_directories(${PROJECT_SOURCE_DIR}/include) include_directories(${PROJECT_SOURCE_DIR}/plugin) # include and link dirs of cuda and tensorrt, you need adapt them if yours are different if(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64") message("embed_platform on") include_directories(/usr/local/cuda/targets/aarch64-linux/include) link_directories(/usr/local/cuda/targets/aarch64-linux/lib) else() message("embed_platform off") # cuda include_directories("C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.8/include") link_directories("C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.8/lib/x64") # tensorrt include_directories("C:/TensorRT/include") link_directories("C:/TensorRT/lib") endif() add_library(myplugins SHARED ${PROJECT_SOURCE_DIR}/plugin/yololayer.cu) target_link_libraries(myplugins nvinfer cudart) set(OpenCV_DIR "C:/opencv/build") find_package(OpenCV) include_directories(${OpenCV_INCLUDE_DIRS}) file(GLOB_RECURSE SRCS ${PROJECT_SOURCE_DIR}/src/*.cpp ${PROJECT_SOURCE_DIR}/src/*.cu) add_executable(yolo11_det ${PROJECT_SOURCE_DIR}/yolo11_det.cpp ${SRCS}) target_link_libraries(yolo11_det nvinfer) target_link_libraries(yolo11_det cudart) target_link_libraries(yolo11_det myplugins) target_link_libraries(yolo11_det ${OpenCV_LIBS}) add_executable(yolo11_cls ${PROJECT_SOURCE_DIR}/yolo11_cls.cpp ${SRCS}) target_link_libraries(yolo11_cls nvinfer) target_link_libraries(yolo11_cls cudart) target_link_libraries(yolo11_cls myplugins) target_link_libraries(yolo11_cls ${OpenCV_LIBS}) add_executable(yolo11_seg ${PROJECT_SOURCE_DIR}/yolo11_seg.cpp ${SRCS}) target_link_libraries(yolo11_seg nvinfer) target_link_libraries(yolo11_seg cudart) target_link_libraries(yolo11_seg myplugins) target_link_libraries(yolo11_seg ${OpenCV_LIBS}) add_executable(yolo11_pose ${PROJECT_SOURCE_DIR}/yolo11_pose.cpp ${SRCS}) target_link_libraries(yolo11_pose nvinfer) target_link_libraries(yolo11_pose cudart) target_link_libraries(yolo11_pose myplugins) target_link_libraries(yolo11_pose ${OpenCV_LIBS})
在训练环境中将pt转onnx，用到了pytorch。
在部署环境中将onnx转engine，不再使用pytorch，用tensorrt自带工具 trtexec.exe 转。位于C:\TensorRT\bin
参考 trtexec命令用法_trtexec 执行-CSDN博客
【一些可能问题】
MATLAB解决cuda编译报错的问题_error: static assertion failed with "error stl1002-CSDN博客
