小白也能学，从0到部署yolov5教程，Windows Linux PC arm Jeston全平台部署-(下)

注意!本篇首发于GiantPandaCV公众号,未经允许,不得转载!
需要的repo: 欢迎star
https://github.com/msnh2012/MsnhnetModelZoo
https://github.com/msnh2012/Msnhnet
https://github.com/msnh2012/CasiaLabeler

接上一篇使用yolov5训练自己的数据集。本篇将分windows和linux, pc和jetson nx平台分别给大家讲解如何使用Msnhnet部署yolov5.

pytorch模型转msnhnet

在yolov5文件夹下打开终端。将best.pt拷贝至weights文件夹下。执行

 python yolov5ToMsnhnet.py 

yolov5ToMsnhnet.py文件内容:

from PytorchToMsnhnet import *
Msnhnet.Export = True
from models.experimental import attempt_load
import torch

weights     = "weights/best.pt" # pt文件
msnhnetPath = "yolov5m.msnhnet" # 导出.msnhnet文件
msnhbinPath = "yolov5m.msnhbin" # 导出.msnhbin文件

model = attempt_load(weights, "cpu") 
model.eval() # cpu模式，推理模式

img = torch.rand(512*512*3).reshape(1,3,512,512) #生成随机推理数据
 
tans(model,img,msnhnetPath,msnhbinPath) #模型转换

导出成功后会在文件夹下生成yolov5m.msnhnet和yolov5m.msnhbin文件。

Windows 篇

1. 准备工作
(1) 安装Visual studio

• 网址:https://visualstudio.microsoft.com/zh-hans/
• 下载visual studio 2017以上任意版本进行安装。此处勾选.Net桌面开发和使用c++的桌面开发。

(2) 安装cuda和cudnn, 此处请自行百度.

• cuda网址:https://developer.nvidia.com/cuda-downloads
• cudnn网址:https://developer.nvidia.com/zh-cn/cudnn
• 下载cudaxx.exe文件安装cuda(此过程最好使用cuda中自带的显卡驱动程序),下载cudnnxxx.zip文件，将其解压到 C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\vxx.xx文件夹下,即完成了cuda和cudnn配置。
• 将C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\vxx.xx 添加到系统环境变量。

(3) 打开cmd, 输入nvcc.测试cuda是否安装完成。以下结果说明cuda配置完成。

(4) 安装cmake(建议3.17).

• cmake下载网址:https://cmake.org/files/v3.17/
• 下载文件: cmake-3.17.5-win64-x64.msi
• 完成安装。

(5) clone Msnhnet
git clone https://github.com/msnh2012/Msnhnet.git
2. 编译OpenCV库
(1) 小编这里给大家准备好了OpenCV的源码文件,不用科学上网了。
链接：https://pan.baidu.com/s/1lpyNNdYqdKj8R-RCQwwCWg 提取码：6agk
(2) 打开cmake-gui.exe。
(3) 点击config选择安装的visual studio版本，选择x64(此处以VS2017为例),点击Finish,等待配置完成.

(4) 参数配置.

  - CMAKE_INSTALL_PREFIX #指定安装位置,如: D:/libs/opencv
  - CPU_BASELINE #选择AVX2(如果CPU支持AVX2加速)
  - BUILD_TESTS #取消勾选

(5) 点击generate->Generating done.
(6) 点击Open Project.分别选择Debug右键生成。(此过程需要等待10min~60min不等,根据电脑配置)

(7) 右键安装。(会将编译好的可执行文件安装在指定安装位置，如:D:/libs/opencv)

(8) 重复6-7步选择Release版本进行编译安装。
(9) 指定OpenCV_DIR环境变量,用于CMakeList能使用FindPackage找到OpenCV.

(10) 指定Path环境变量.在Path环境变量下添加Opencv的bin文件夹位置,如: D:\libs\opencv\x64\vc15\bin

3. 编译Msnhnet库
(1) 打开cmake-gui.exe。

(2) 点击config选择安装的visual studio版本，选择x64(此处以VS2017为例),点击Finish,等待配置完成.

(3) 勾选以下参数。

 - CMAKE_INSTALL_PREFIX #指定安装位置,如: D:/libs/Msnhnet
 - BUILD_EXAMPLE      #构建示例
 - BUILD_SHARED_LIBS  #构建动态链接库
 - BUILD_USE_CUDNN    #使用CUDNN
 - BUILD_USE_GPU      #使用GPU
 - BUILD_USE_OPENCV   #使用OPENCV
 - ENABLE_OMP         #使用OMP
 - OMP_MAX_THREAD     #使用最大核心数

(4) 点击generate->Generating done.
(5) 点击Open Project.分别选择Debug右键生成。

(6) 右键安装。(会将编译好的可执行文件安装在指定安装位置，如:D:/libs/Msnhnet)

(7) 重复6-7步选择Release版本进行编译安装.
(8) 指定Msnhnet_DIR环境变量,用于CMakeList能使用FindPackage找到Msnhnet.

(9)指定Path环境变量.在Path环境变量下添加Msnhnet的bin文件夹位置,如: D:\libs\Msnhnet\bin
(10) 测试。

• 下载小编准备好的一系列Msnhnet测试模型。并解压到如D盘根目录
  链接：https://pan.baidu.com/s/1mBaJvGx7tp2ZsLKzT5ifOg
  提取码：x53z
• 在Msnhnet安装目录打开终端。执行

yolov3_gpu D:/models
yolov3_gpu_fp16 D:/models #fp16推理

• 当然,你可以可以测试其它模型。

4. 使用C#部署Msnhnet
(1) clone MsnhnetSharp

git clone https://github.com/msnh2012/MsnhnetSharp

(2) 双击打开MsnhnetSharp.sln文件
MsnhnetSharp源码:

using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;
using System.Runtime.InteropServices;
using System.Drawing;
using System.Drawing.Imaging;
using static MsnhnetSharp.MsnhnetDef;

namespace MsnhnetSharp
{
    public class Msnhnet
    {
        const string MsnhnetLib = "msnhnet.dll";

        private const int MaxBBoxNum = 1024;

        [DllImport(MsnhnetLib, EntryPoint = "initMsnhnet")]
        static extern int _initMsnhnet();

        [DllImport(MsnhnetLib, EntryPoint = "dispose")]
        static extern int _dispose();

        [DllImport(MsnhnetLib, EntryPoint = "withGPU")]
        static extern int _withGPU(ref int GPU);

        [DllImport(MsnhnetLib, EntryPoint = "withCUDNN")]
        static extern int _withCUDNN(ref int CUDNN);

        [DllImport(MsnhnetLib, EntryPoint = "getInputDim")]
        static extern int _getInputDim(ref int width, ref int heigth, ref int channel);

        [DllImport(MsnhnetLib, EntryPoint = "getCpuForwardTime")]
        static extern int _getCpuForwardTime(ref float cpuForwardTime);

        [DllImport(MsnhnetLib, EntryPoint = "getGpuForwardTime")]
        static extern int _getGpuForwardTime(ref float getGpuForwardTime);

        [DllImport(MsnhnetLib, EntryPoint = "buildMsnhnet")]
        static extern int _buildMsnhnet(ref IntPtr msg, string msnhnet, string msnhbin, int useFp16, int useCudaOnly);

        [DllImport(MsnhnetLib, EntryPoint = "runClassifyFile")]
        static unsafe extern int _runClassifyFile(ref IntPtr msg, string imagePath, ref int bestIndex, PredDataType predDataType,
                                                 int runGPU, float* mean, float* std);

        [DllImport(MsnhnetLib, EntryPoint = "runClassifyList")]
        static unsafe extern int _runClassifyList(ref IntPtr msg, byte* data, int width, int height, int channel, ref int bestIndex, PredDataType predDataType,
                                                  int swapRGB, int runGPU, float* mean, float* std);

        [DllImport(MsnhnetLib, EntryPoint = "runClassifyNoPred")]
        static unsafe extern int _runClassifyNoPred(ref IntPtr msg, float* data, int len, ref int bestIndex, int runGPU);

        [StructLayout(LayoutKind.Sequential)]
        public struct BBox
        {
            public float x;
            public float y;
            public float w;
            public float h;
            public float conf;
            public float bestClsConf;
            public UInt32 bestClsIdx;
            public float angle;
        };

        [StructLayout(LayoutKind.Sequential)]
        public struct BBoxContainer
        {
            [MarshalAs(UnmanagedType.ByValArray, SizeConst = MaxBBoxNum)]
            public BBox[] boxes;
        }

        [DllImport(MsnhnetLib, EntryPoint = "runYoloFile")]
        static extern int _runYoloFile(ref IntPtr msg, string imagePath, ref BBoxContainer bboxContainer, ref int detectedNum, int runGPU);

        [DllImport(MsnhnetLib, EntryPoint = "runYoloList")]
        static unsafe extern int _runYoloList(ref IntPtr msg, byte* data, int width, int height, int channel, ref BBoxContainer bboxContainer, ref int detectedNum, int swapRGB, int runGPU);

        private bool netBuilt = false;
        private bool netInited = false;

        /// 

        /// check GPU 
        /// 
        /// 
        static public bool WithGPU()
        {
            int GPU = 0;
            _withGPU(ref GPU);
            return (GPU == 1) ? true : false;
        }

        /// 

        /// check CUDNN
        /// 
        /// 
        static public bool WithCudnn()
        {
            int CUDNN = 0;
            _withCUDNN(ref CUDNN);
            return (CUDNN == 1) ? true : false;
        }

        /// 

        /// Get input dim
        /// 
        /// 
        public Dim GetInputDim()
        {
            if (netBuilt)
            {
                int width = 0;
                int height = 0;
                int channel = 0;
                _getInputDim(ref width, ref height, ref channel);

                Dim dim;
                dim.width = width;
                dim.height = height;
                dim.channel = channel;

                return dim;
            }
            else
            {
                throw new Exception("Net wasn't built yet");
            }

        }

        /// 

        /// Get cpu forward time
        /// 
        /// 
        public float GetCpuForwardTime()
        {
            float time = 0;
            _getCpuForwardTime(ref time);
            return time;
        }

        /// 

        /// Get gpu forward time
        /// 
        /// 
        public float GetGpuForwardTime()
        {
            float time = 0;
            if (_getGpuForwardTime(ref time) != 1)
            {
                throw new Exception("Msnhnet is not compiled with GPU mode!");
            }
            return time;
        }

        /// 

        /// init net
        /// 
        public void InitNet()
        {
            if (_initMsnhnet() != 1)
            {
                throw new Exception("Init net Failed");
            }
            netInited = true;
        }

        /// 

        /// build net
        /// 
        /// msnhnet file path
        /// msnhbin file path
        /// use fp16 or not
        /// build with cudnn, but only use cuda
        public void BuildNet(string msnhnet, string msnhbin, bool useFp16, bool useCudaOnly)
        {
            if (!netInited)
            {
                throw new Exception("Net wasn't inited yet");
            }

            IntPtr msg = new IntPtr();
            if (_buildMsnhnet(ref msg, msnhnet, msnhbin, useFp16 ? 1 : 0, useCudaOnly ? 1 : 0) != 1)
            {
                string mstr = Marshal.PtrToStringAnsi(msg);
                throw new Exception(mstr);
            }
            netBuilt = true;
        }

        /// 

        /// dispose net
        /// 
        public void Dispose()
        {
            _dispose();
        }

        /// 

        /// forward net with image file, with preprocess
        /// 
        /// image file path
        /// process function
        /// run wit GPU
        ///  if normalize mean val 
        /// if normalize std val
        /// 
        public int RunClassifyFile(string imagePath, PredDataType predDataType, bool runGPU, float[] mean = null, float[] std = null)
        {
            if (!netBuilt)
            {
                throw new Exception("Net wasn't built yet");
            }

            IntPtr msg = new IntPtr();
            int bestIndex = 0;
            unsafe
            {
                if (predDataType == PredDataType.PRE_DATA_TRANSFORMED_FC3)
                {
                    fixed (float* meanPtr = mean)
                    {
                        fixed (float* stdPtr = std)
                        {
                            if (_runClassifyFile(ref msg, imagePath, ref bestIndex, predDataType, runGPU ? 1 : 0, meanPtr, stdPtr) != 1)
                            {
                                string mstr = Marshal.PtrToStringAnsi(msg);
                                throw new Exception(mstr);
                            }
                        }

                    }
                }
                else
                {
                    if (_runClassifyFile(ref msg, imagePath, ref bestIndex, predDataType, runGPU ? 1 : 0, null, null) != 1)
                    {
                        string mstr = Marshal.PtrToStringAnsi(msg);
                        throw new Exception(mstr);
                    }
                }
            }

            return bestIndex;
        }

        /// 

        ///  forward net with image BitmapData, with preprocess
        /// 
        /// data
        /// process function
        /// net swap RGB or not
        /// run wit GPU
        ///  if normalize mean val 
        /// if normalize std val
        /// 
        public int RunClassifyList(BitmapData bitmap, PredDataType predDataType, bool swapRGB, bool runGPU, float[] mean = null, float[] std = null)
        {
            if (!netBuilt)
            {
                throw new Exception("Net wasn't built yet");
            }
            IntPtr msg = new IntPtr();
            int bestIndex = 0;
            unsafe
            {
                if (predDataType == PredDataType.PRE_DATA_TRANSFORMED_FC3)
                {
                    fixed (float* meanPtr = mean)
                    {
                        fixed (float* stdPtr = std)
                        {
                            if (_runClassifyList(ref msg, (byte*)bitmap.Scan0, bitmap.Width, bitmap.Height, bitmap.Stride / bitmap.Width, ref bestIndex, predDataType, swapRGB ? 1 : 0, runGPU ? 1 : 0, meanPtr, stdPtr) != 1)
                            {
                                string mstr = Marshal.PtrToStringAnsi(msg);
                                throw new Exception(mstr);
                            }
                        }

                    }
                }
                else
                {
                    if (_runClassifyList(ref msg, (byte*)bitmap.Scan0, bitmap.Width, bitmap.Height, bitmap.Stride / bitmap.Width, ref bestIndex, predDataType, swapRGB ? 1 : 0, runGPU ? 1 : 0, null, null) != 1)
                    {
                        string mstr = Marshal.PtrToStringAnsi(msg);
                        throw new Exception(mstr);
                    }
                }
            }

            return bestIndex;
        }

        /// 

        /// forward net with float data, without preprocess
        /// 
        /// flaot data
        /// run wit GPU
        /// 
        public int RunClassifyNoPred(float[] data, bool runGPU)
        {
            if (!netBuilt)
            {
                throw new Exception("Net wasn't built yet");
            }
            IntPtr msg = new IntPtr();
            int bestIndex = 0;
            unsafe
            {
                fixed (float* dataPtr = data)
                {
                    if (_runClassifyNoPred(ref msg, dataPtr, data.Length, ref bestIndex, runGPU ? 1 : 0) != 1)
                    {
                        string mstr = Marshal.PtrToStringAnsi(msg);
                        throw new Exception(mstr);
                    }
                }
            }

            return bestIndex;
        }

        /// 

        /// run yolo
        /// 
        /// image path
        /// run with GPU
        /// 
        public List RunYoloFile(string imagePath, bool runGPU)
        {
            List bboxVec = new List();
            IntPtr msg = new IntPtr();
            BBoxContainer bboxContainer = new BBoxContainer();
            int detectNum = 0;

            if (!netBuilt)
            {
                throw new Exception("Net wasn't built yet");
            }

            if (_runYoloFile(ref msg, imagePath, ref bboxContainer, ref detectNum, runGPU?1:0)!=1)
            {
                string mstr = Marshal.PtrToStringAnsi(msg);
                throw new Exception(mstr);
            }

            for (int i = 0; i < detectNum; i++)
            {
                bboxVec.Add(bboxContainer.boxes[i]);
            }

            return bboxVec;
        }

        public List RunYoloList(BitmapData bitmap, bool swapRGB, bool runGPU)
        {
            if (!netBuilt)
            {
                throw new Exception("Net wasn't built yet");
            }
            List bboxVec = new List();
            IntPtr msg = new IntPtr();
            BBoxContainer bboxContainer = new BBoxContainer();
            int detectNum = 0;

            unsafe
            {
                if (_runYoloList(ref msg, (byte*)bitmap.Scan0, bitmap.Width, bitmap.Height, bitmap.Stride / bitmap.Width, ref bboxContainer, ref detectNum, swapRGB?1:0, runGPU?1:0) != 1)
                {
                    string mstr = Marshal.PtrToStringAnsi(msg);
                    throw new Exception(mstr);
                }
            }

            for (int i = 0; i < detectNum; i++)
            {
                bboxVec.Add(bboxContainer.boxes[i]);
            }

            return bboxVec;
        }
    }
}

(3) 选择x64平台和Release模式,右键生成MsnhnetSharp,再生成MsnhnetForm.
(4) 点击启动按钮。
(5) 在参数配置栏,分别指定msnhnetPath和msnhbinPath为之前导出的yolov5m的参数。然后将上一篇制作好的labels.txt文件,复制一份,重命名为labels.names.
(6) 点击初始化网络。等待初始化完成，init done.
(7) 点击读取图片, 选择那张bus.jpg.
(8) 点击yolo GPU(Yolo Detect GPU). 第一次推理时间较长。
(9) 点击重置图片。
(10) 再次点击yolo GPU(Yolo Detect GPU). 随后推理时间正常.

至此,使用C#部署Msnhnet完成,后续可以参考MsnhnetForm将MsnhnetSharp部署到你自己的工程中。

5. 使用CMake部署Msnhnet
工程文件源码:
链接：https://pan.baidu.com/s/1lpyNNdYqdKj8R-RCQwwCWg 提取码：6agk
(1) 新建MsnhnetPrj文件夹
(2) 将yolov5m.msnhnet,yolov5m.msnhbin,labels.txt拷贝到MsnhnetPrj文件夹内
(3) 新建CMakeLists.txt文件

cmake_minimum_required(VERSION 3.15)
project(yolov5m_msnhnet
        LANGUAGES CXX C CUDA
        VERSION 1.0)

find_package(OpenCV REQUIRED)
find_package(Msnhnet REQUIRED)
find_package(OpenMP REQUIRED)

add_executable(yolov5m_msnhnet yolov5m_msnhnet.cpp)

target_include_directories(yolov5m_msnhnet PUBLIC ${Msnhnet_INCLUDE_DIR})
target_link_libraries(yolov5m_msnhnet PUBLIC ${OpenCV_LIBS} Msnhnet)

(4) 新建yolov5m_msnhnet.cpp文件

#include 
#include "Msnhnet/net/MsnhNetBuilder.h"
#include "Msnhnet/io/MsnhIO.h"
#include "Msnhnet/config/MsnhnetCfg.h"
#include "Msnhnet/utils/MsnhOpencvUtil.h"

void yolov5sGPUOpencv(const std::string& msnhnetPath, const std::string& msnhbinPath, const std::string& imgPath,  const std::string& labelsPath)
{
    try
    {
        Msnhnet::NetBuilder  msnhNet;
        Msnhnet::NetBuilder::setOnlyGpu(true);
        //msnhNet.setUseFp16(true); //开启使用FP16推理
        msnhNet.buildNetFromMsnhNet(msnhnetPath);
        std::cout<<msnhNet.getLayerDetail();
        msnhNet.loadWeightsFromMsnhBin(msnhbinPath);
        std::vector<std::string> labels ;
        Msnhnet::IO::readVectorStr(labels, labelsPath.data(), "\n");
        Msnhnet::Point2I inSize = msnhNet.getInputSize();

        std::vector<float> img;
        std::vector<std::vector<Msnhnet::YoloBox>> result;

        img = Msnhnet::OpencvUtil::getPaddingZeroF32C3(imgPath, cv::Size(inSize.x,inSize.y));
        for (size_t i = 0; i < 10; i++)
        {
            auto st = Msnhnet::TimeUtil::startRecord();
            result = msnhNet.runYoloGPU(img);
            std::cout<<"time  : " << Msnhnet::TimeUtil::getElapsedTime(st) <<"ms"<<std::endl<<std::flush;
        }

        cv::Mat org = cv::imread(imgPath);
        Msnhnet::OpencvUtil::drawYoloBox(org,labels,result,inSize);
        cv::imshow("test",org);
        cv::waitKey();
    }
    catch (Msnhnet::Exception ex)
    {
        std::cout<<ex.what()<<" "<<ex.getErrFile() << " " <<ex.getErrLine()<< " "<<ex.getErrFun()<<std::endl;
    }
}

int main(int argc, char** argv)
{
    std::string msnhnetPath = "yolov5m.msnhnet";
    std::string msnhbinPath = "yolov5m.msnhbin";
    std::string labelsPath  = "labels.txt";
    std::string imgPath = "bus.jpg";

    yolov5sGPUOpencv(msnhnetPath, msnhbinPath, imgPath,labelsPath);
    getchar();

    return 0;
}

(5) 配置CMake
打开cmake-gui.exe,按以下配置.点击Config.Generate

(6) 编译,点击Open Project,选择Release模式,参靠之前编译Msnhnet直接生成。

(7) 拷贝可执行文件。
从MsnhnetPrj/build/Release/yolov5m_msnhnet.exe拷贝到MsnhnetPrj目录。
(8) 部署结果
双击yolov5m_msnhnet.exe查看部署结果

Linux(Jetson NX) 篇

1. 准备工作
一般来说，Jetson都已经自带了cuda和cudnn,故不用专门安装。

• 安装构建工具

sudo apt-get install build-essential

• 安装opencv

sudo apt-get install libopencv 

2. 编译Msnhnet库
(1) 终端打开cmake-gui。

(2) 点击config选择安装的visual studio版本，选择x64(此处以VS2017为例),点击Finish,等待配置完成.

(3) 勾选以下参数。

 - CMAKE_INSTALL_PREFIX #指定安装位置,如: D:/libs/Msnhnet
 - BUILD_EXAMPLE      #构建示例
 - BUILD_SHARED_LIBS  #构建动态链接库
 - BUILD_USE_CUDNN    #使用CUDNN
 - BUILD_USE_GPU      #使用GPU
 - BUILD_USE_NEON     #使用neon加速
 - BUILD_USE_OPENCV   #使用OPENCV
 - ENABLE_OMP         #使用OMP
 - OMP_MAX_THREAD     #使用最大核心数

(4) 点击generate->Generating done.
(5) 在Msnhnet/build文件夹中打开终端。

make -j
sudo make install

(7) 配置系统环境变量

sudo gedit /etc/ld.so.conf.d/usr.confg
# 添加: /usr/local/lib
sudo ldconfig

(8) 测试。

• 下载小编准备好的一系列Msnhnet测试模型。并解压到如home根目录
  链接：https://pan.baidu.com/s/1mBaJvGx7tp2ZsLKzT5ifOg
  提取码：x53z

cd /usr/local/bin
yolov3_gpu /home/xxx/models
yolov3_gpu_fp16 /home/xxx/models #fp16推理

• 当然,你可以可以测试其它模型。

3. 使用CMake部署Msnhnet
工程文件源码:
链接：https://pan.baidu.com/s/1lpyNNdYqdKj8R-RCQwwCWg 提取码：6agk

(1) 新建MsnhnetPrj文件夹
(2) 将yolov5m.msnhnet,yolov5m.msnhbin，labels.txt拷贝到MsnhnetPrj文件夹内
(3) 新建CMakeLists.txt文件

cmake_minimum_required(VERSION 3.15)
project(yolov5m_msnhnet
        LANGUAGES CXX C CUDA
        VERSION 1.0)

find_package(OpenCV REQUIRED)
find_package(Msnhnet REQUIRED)
find_package(OpenMP REQUIRED)

add_executable(yolov5m_msnhnet yolov5m_msnhnet.cpp)

target_include_directories(yolov5m_msnhnet PUBLIC ${Msnhnet_INCLUDE_DIR})
target_link_libraries(yolov5m_msnhnet PUBLIC ${OpenCV_LIBS} Msnhnet)

(4) 新建yolov5m_msnhnet.cpp文件

#include 
#include "Msnhnet/net/MsnhNetBuilder.h"
#include "Msnhnet/io/MsnhIO.h"
#include "Msnhnet/config/MsnhnetCfg.h"
#include "Msnhnet/utils/MsnhOpencvUtil.h"


void yolov5sGPUOpencv(const std::string& msnhnetPath, const std::string& msnhbinPath, const std::string& imgPath,  const std::string& labelsPath)
{
    try
    {
        Msnhnet::NetBuilder  msnhNet;
        Msnhnet::NetBuilder::setOnlyGpu(true);
        //msnhNet.setUseFp16(true); //开启使用FP16推理
        msnhNet.buildNetFromMsnhNet(msnhnetPath);
        std::cout<<msnhNet.getLayerDetail();
        msnhNet.loadWeightsFromMsnhBin(msnhbinPath);
        std::vector<std::string> labels ;
        Msnhnet::IO::readVectorStr(labels, labelsPath.data(), "\n");
        Msnhnet::Point2I inSize = msnhNet.getInputSize();

        std::vector<float> img;
        std::vector<std::vector<Msnhnet::YoloBox>> result;

        img = Msnhnet::OpencvUtil::getPaddingZeroF32C3(imgPath, cv::Size(inSize.x,inSize.y));
        for (size_t i = 0; i < 10; i++)
        {
            auto st = Msnhnet::TimeUtil::startRecord();
            result = msnhNet.runYoloGPU(img);
            std::cout<<"time  : " << Msnhnet::TimeUtil::getElapsedTime(st) <<"ms"<<std::endl<<std::flush;
        }

        cv::Mat org = cv::imread(imgPath);
        Msnhnet::OpencvUtil::drawYoloBox(org,labels,result,inSize);
        cv::imshow("test",org);
        cv::waitKey();
    }
    catch (Msnhnet::Exception ex)
    {
        std::cout<<ex.what()<<" "<<ex.getErrFile() << " " <<ex.getErrLine()<< " "<<ex.getErrFun()<<std::endl;
    }
}

int main(int argc, char** argv)
{
    std::string msnhnetPath = "../yolov5m.msnhnet";
    std::string msnhbinPath = "../yolov5m.msnhbin";
    std::string labelsPath  = "../labels.txt";
    std::string imgPath = "../bus.jpg";

    yolov5sGPUOpencv(msnhnetPath, msnhbinPath, imgPath,labelsPath);
    getchar();

    return 0;
}

(5) 编译,在MsnhnetPrj文件夹下打开终端

mkdir build
cd build 
make 
./yolov5m_msnhnet

(6) 部署结果

Linux(PC) 篇

和Jetson Nx部署类似, 主要区别是先要在Linux上配置好cuda和cudnn, 然后卸载CMake, 安装CMake 3.17版本. 其它的和Jestson NX一样.(ps. 在CMake参数配置里没有NEON项,此为ARM平台专有)

到此,使用Msnhnet从0到部署Yolov5网络完成。

最后

• 欢迎关注我和Buff及公众号的小伙伴们一块维护的一个深度学习框架Msnhnet:
• https://github.com/msnh2012/Msnhnet