当前位置：首页 > news >正文

YoloV10 训练自己的数据集（推理，转化，C#部署）

news 2025/7/12 10:10:09

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

也可以看我另一篇文章

Yolov10训练，转化onnx，推理_yolov10转onnx-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 simplify

detect.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);}}
}