目录
一、下载
三、开始训练
train.py
detect.py
export.py
超参数都在这个路径下
四、C#读取yolov10模型进行部署推理
如下程序是用来配置openvino
配置好引用后就可以生成dll了 再创建一个控件，作为显示 net framework 4.8版本的
再nuget工具箱里下载 opencvsharp4 以及openvino
然后主流程代码
效果
我的yolov10 训练源码
C#部署yolov10源码
一、下载
GitHub - THU-MIG/yolov10: YOLOv10: Real-Time End-to-End Object Detection
或者你可以再浏览器搜索框里直接搜索 yolov10 github
二、环境配置
下载anaconda 并安装 在网上随意下载一个2022版本的就行
yolov10和yolov8的文件结构差不多 所以如果你训练过其他的yolov5以上的yolo，你可以直接拷贝环境进行使用，当然你如果想配置gpu
就需要cuda cudnn 和 gpu版本的torch
其他的直接pip install 即可
pip install requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
下方网站中下载你需要的版本下载时要注意对应关系，
cuda和cudnn各个版本的Pytorch下载网页版，onnx，ncnn，pt模型转化工具_cuda国内镜像下载网站-CSDN博客
三、开始训练
有一点要注意v10版本其实是从v8版本上面改的 所以v10的预训练模型你需要自行下载 否则就会下载成v8的
首先标注数据集 在pycharm 中下载labelimg
pip install labelimg -i https://pypi.tuna.tsinghua.edu.cn/simple
下载好后直接在终端输入labelimg 开始标注 训练流程基本和yolov5差不多
在yolov10的根目录下创建一个名为data的文件夹 里面再创建一个data.yaml文件 用于数据集的读取
再在根目录创建三个py文件 分别是 train.py detect.py export.py
train.py
from ultralytics import YOLOv10model_yaml_path = "ultralytics/cfg/models/v10/yolov10s.yaml"
#数据集配置文件
data_yaml_path = 'data/data.yaml'
#预训练模型
pre_model_name = 'yolov10s.pt'if __name__ == '__main__':#加载预训练模型model = YOLOv10(model_yaml_path).load(pre_model_name)#训练模型results = model.train(data=data_yaml_path,epochs=450,batch=8,device=0,name='train/exp')# yolo export model="H:\\DL\\yolov10-main\\runs\\detect\\train\\exp\\weights\\best.pt" format=onnx opset=13 simplifydetect.py
from ultralytics import YOLOv10import torch
if torch.cuda.is_available():device = torch.device("cuda")
else:raise Exception("CUDA is not")model_path = r"H:\\DL\\yolov10-main\\runs\\detect\\train\\exp4\\weights\\best.pt"
model = YOLOv10(model_path)
results = model(source=r'H:\DL\yolov10-main\dataDakeset\two_CD_double\test',name='predict/exp',conf=0.45,save=True,device='0')
export.py
from ultralytics import YOLOv10
model=YOLOv10("H:\\DL\\yolov10-main\\runs\\detect\\train\\exp\\weights\\best.pt")model.export(format='onnx')# 'torchscript, onnx, openvino, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs, paddle'
超参数都在这个路径下
然后设置好参数就可以直接训练了
推理用detect.py 推理 转化用export.py 转化， 转化为哪种模型 就替换即可
四、C#读取yolov10模型进行部署推理
我们需要设定yolov10的模型结构
using OpenCvSharp;
using OpenVinoSharp.Extensions.result;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;namespace Yolov10_DLLnet
{public class YOLOv10Det : YOLO{public YOLOv10Det(string model_path, string engine, string device, int categ_nums, float det_thresh, float det_nms_thresh, int input_size): base(model_path, engine, device, categ_nums, det_thresh, det_nms_thresh, new int[] { 1, 3, input_size, input_size },new List<string> { "images" }, new List<int[]> { new int[] { 1, 4 + categ_nums, 8400 } }, new List<string> { "output0" }){}protected override BaseResult postprocess(List<float[]> results){List<Rect> positionBoxes = new List<Rect>();List<int> classIds = new List<int>();List<float> confidences = new List<float>();// Preprocessing output resultsfor (int i = 0; i < results[0].Length / 6; i++){int s = 6 * i;if ((float)results[0][s + 4] > 0.5){float cx = results[0][s + 0];float cy = results[0][s + 1];float dx = results[0][s + 2];float dy = results[0][s + 3];int x = (int)((cx) * m_factor);int y = (int)((cy) * m_factor);int width = (int)((dx - cx) * m_factor);int height = (int)((dy - cy) * m_factor);Rect box = new Rect();box.X = x;box.Y = y;box.Width = width;box.Height = height;positionBoxes.Add(box);classIds.Add((int)results[0][s + 5]);confidences.Add((float)results[0][s + 4]);}}DetResult re = new DetResult();// for (int i = 0; i < positionBoxes.Count; i++){re.add(classIds[i], confidences[i], positionBoxes[i]);}return re;}}
}
然后再设置各项参数 你可以再其中自己定义一个文件 里面写上你需要的类 比如置信度，类别 以及类别数量等等。为后续的dll生成做准备。
如下程序是用来配置openvino
using OpenCvSharp.Dnn;
using OpenCvSharp;
using OpenVinoSharp;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Runtime.InteropServices;
using System.Text;
using System.Threading.Tasks;
using System.Reflection;
//using static System.Windows.Forms.Design.AxImporter;namespace Yolov10_DLLnet
{public class Predictor : IDisposable{private Core core;private Model model;private CompiledModel compiled;private InferRequest openvino_infer;private Net opencv_infer;private string engine = null;public Predictor() { }public Predictor(string model_path, string engine, string device){if (model_path == null){throw new ArgumentNullException(nameof(model_path));}this.engine = engine;if (engine == "OpenVINO"){core = new Core();model = core.read_model(model_path);compiled = core.compile_model(model, device);openvino_infer = compiled.create_infer_request();}}public void Dispose(){openvino_infer.Dispose();compiled.Dispose();model.Dispose();core.Dispose();GC.Collect();}public List<float[]> infer(float[] input_data, List<string> input_names, int[] input_size, List<string> output_names, List<int[]> output_sizes){List<float[]> returns = new List<float[]>();var input_tensor = openvino_infer.get_input_tensor();input_tensor.set_data(input_data);openvino_infer.infer();foreach (var name in output_names){var output_tensor = openvino_infer.get_tensor(name);returns.Add(output_tensor.get_data<float>((int)output_tensor.get_size()));}return returns;}}
}
创建一个名为yolo的cs文件用于 将yolov10模型结构做引用
//using Microsoft.VisualBasic.Logging;
using OpenCvSharp;
using OpenVinoSharp.Extensions.model;
using OpenVinoSharp.Extensions.process;
using OpenVinoSharp.Extensions.result;
using System;
using System.Collections.Generic;
using System.Diagnostics;
using System.Linq;
using System.Text;
using System.Threading.Tasks;
//using Yolov10_DLLnet;
using static OpenVinoSharp.Node;namespace Yolov10_DLLnet
{public class YOLO : IDisposable{protected int m_categ_nums;protected float m_det_thresh;protected float m_det_nms_thresh;protected float m_factor;protected int[] m_input_size;protected List<int[]> m_output_sizes;protected List<string> m_input_names;protected List<string> m_output_names;protected List<int> m_image_size = new List<int>();private Predictor m_predictor;Stopwatch sw = new Stopwatch();public YOLO(){m_predictor = new Predictor();}public YOLO(string model_path, string engine, string device, int categ_nums, float det_thresh,float det_nms_thresh, int[] input_size, List<string> input_names, List<int[]> output_sizes, List<string> output_names){m_predictor = new Predictor(model_path, engine, device);m_categ_nums = categ_nums;m_det_thresh = det_thresh;m_det_nms_thresh = det_nms_thresh;m_input_size = input_size;m_output_sizes = output_sizes;m_input_names = input_names;m_output_names = output_names;}float[] preprocess(Mat img){m_image_size = new List<int> { (int)img.Size().Width, (int)img.Size().Height };Mat mat = new Mat();Cv2.CvtColor(img, mat, ColorConversionCodes.BGR2RGB);mat = Resize.letterbox_img(mat, (int)m_input_size[2], out m_factor);mat = Normalize.run(mat, true);return Permute.run(mat);}List<float[]> infer(Mat img){List<float[]> re;float[] data = preprocess(img);re = m_predictor.infer(data, m_input_names, m_input_size, m_output_names, m_output_sizes);return re;}public BaseResult predict(Mat img){List<float[]> result_data = infer(img);BaseResult re = postprocess(result_data);return re;}protected virtual BaseResult postprocess(List<float[]> results){return new BaseResult();}public void Dispose(){m_predictor.Dispose();}public static YOLO GetYolo(string model_type, string model_path, string engine, string device,int categ_nums, float det_thresh, float det_nms_thresh, int input_size){return new YOLOv10Det(model_path, engine, device, categ_nums, det_thresh, det_nms_thresh, input_size);}protected static float sigmoid(float a){float b = 1.0f / (1.0f + (float)Math.Exp(-a));return b;}}
}
配置好引用后就可以生成dll了 再创建一个控件，作为显示 net framework 4.8版本的
再nuget工具箱里下载 opencvsharp4 以及openvino
然后主流程代码
//using Microsoft.VisualBasic.Logging;
using OpenCvSharp;
using OpenVinoSharp.Extensions.process;
using OpenVinoSharp.Extensions.result;
using OpenVinoSharp.Extensions.utility;
using SharpCompress.Common;
using System.Collections.Generic;
using System;
using System.Diagnostics;
using System.Runtime.CompilerServices;
using System.Windows.Forms;
using static OpenVinoSharp.Node;
using static System.Windows.Forms.VisualStyles.VisualStyleElement;
using Point = OpenCvSharp.Point;using Yolov10_DLLnet;
using System.Drawing;
using ZstdSharp.Unsafe;namespace YOLOV10_WinformDemo
{public partial class Form1 : Form{//string filePath = "";private YOLO yolo;public Form1(){InitializeComponent();yolo = new YOLO();//string model_path = "best_0613.onnx";}/// <summary>/// yolov10 onnx模型文件路径/// </summary>private string model_path = "H:\\YCDandPCB_Yolov5_net\\Yolov10_and_Yolov5Seg\\yolov10_Detztest\\YOLOV10_WinformDemo\\bestV10det.onnx";/// <summary>/// 开始识别/// </summary>/// <param name="sender"></param>/// <param name="e"></param>private void button2_Click(object sender, EventArgs e){Stopwatch sw = new Stopwatch();OpenFileDialog openFile = new OpenFileDialog();string filePath = "";if (openFile.ShowDialog() == DialogResult.OK){filePath = openFile.FileName;}//目标检测//string model_path = "best_0613.onnx";classesLabel label = new classesLabel();string model_type = "YOLOv10Det";string engine_type = "OpenVINO";string device = "CPU";//################# 阈值 #######################################float score = label.Score_Threshold;float nms = label.NMS_Threshold;int categ_num = label.classes_count_1;int input_size = label.W_H_size_1;yolo = YOLO.GetYolo(model_type, model_path, engine_type, device, categ_num, score, nms, input_size);//##################### 图片推理阶段 #######################//System.Drawing.Image image = Image.FromFile(openFile.FileName);//string input_path = openFile;Mat img = Cv2.ImRead(filePath);pictureBox1.Image = OpenCvSharp.Extensions.BitmapConverter.ToBitmap(img);sw.Restart();(Mat, BaseResult) re_img = image_predict(img);sw.Stop();pictureBox2.Image = OpenCvSharp.Extensions.BitmapConverter.ToBitmap(re_img.Item1);DetResult detResult = re_img.Item2 as DetResult;for (int i = 0; i < detResult.count; i++){//textBox1.Text = detResult.datas[i].lable;//置信度//textBox2.Text = detResult.datas[i].score.ToString("0.00");int X = detResult.datas[i].box.TopLeft.X;int Y = detResult.datas[i].box.TopLeft.Y;int W = detResult.datas[i].box.Width;int H = detResult.datas[i].box.Height;//textBox3.Text = X.ToString();//textBox4.Text = Y.ToString();//textBox5.Text = W.ToString();//textBox6.Text = H.ToString();Console.WriteLine(X);Console.WriteLine(Y);}// 获取并打印运行时间//TimeSpan ts = sw.Elapsed;textBox7.Text = sw.ElapsedMilliseconds.ToString();}(Mat, BaseResult) image_predict(Mat img, bool is_video = false){Mat re_img = new Mat();//############################# classes ###################################classesLabel label = new classesLabel();List<string> class_names = label.class_names;//开始识别，并返回识别结果BaseResult result = yolo.predict(img);result.update_lable(class_names);re_img = Visualize.draw_det_result(result, img);return (re_img, result);}}
}
效果
我的yolov10 训练源码
【免费】yolov10优化代码，包含，train.py,detect.py,export.py脚本以及预训练模型资源-CSDN文库
C#部署yolov10源码
C#部署YoloV10目标检测.netframework4.8,打开即用内有（主程序和dll生成程序）资源-CSDN文库
