devbox
很楠不爱3
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

yolov7模型部署——代码实现（python版和c#版）_yolov7代码

作者：很楠不爱3 | 2024-03-23 14:10:14

赞

踩

yolov7代码

忙了两个星期，终于把c#版onnx调用整合到项目中，并和UI功能结合起来了~~~也终于腾出时间来总结一下，都快忘记踩过什么坑了T_T。

一，python版

        python版比较容易，毕竟有官方的detect.py指导，据说之前官方放了个使用onnx推理的ipynb文件，但很快就删了~~~参考这篇博客。通过模型结构图查看模型输入要求尺寸为[1 3 640 640]。

onnx调用的基本思想步骤就是：

  1. 根据EP建立session： 
    1. providers = ['CUDAExecutionProvider', 'CPUExecutionProvider'] if cuda else ['CPUExecutionProvider']
    2. session = ort.InferenceSession(w, providers=providers)
  2. 调用session.run: 
    outputs = session.run(outname, inp)[0]
  3. 解析并可视化结果

具体实现步骤为：

  • Resize原始图片，处理成onnx模型需要的大小，这个过程是通过letterbox()函数实现的。公主的模型输入尺寸为[640, 640]。根据官方提供的思想resize的时候需要保持原始图片的长宽比。因此需要首先计算与所需尺寸相差的最小ratio（即根据原始长度长，宽较长一边计算出的ratio），将图片resize之后，剩下的空隙用灰色来填充。公主输入图片原始尺寸为[512, 555, 3],转换之后为[640, 640, 3]，如下图所示：

  • 将图像转为RGB格式，并增加一维以符合输入尺寸，通过以下命令实现 
    image = image.transpose((2, 0, 1))
image = np.expand_dims(image, 0)
  • 调用onnx模型
  • 可视化输出结果，输出是个二维n*7数组，n代表检测结果数，每一条结果的7个值分别代表：[batch_id, bbox左上角横坐标，bbox左上角纵坐标，bbox右下角横坐标，bbox右下角纵坐标，classIndex，置信度]。这里可视化bbox时注意，得出的bbox坐标为resize之后的图像中坐标，需要根据resize时用到的ratio和padding数来还原出原始坐标，才可以在原始图像中显示出来。

python代码：

  1. import cv2
  2. import time
  3. import requests
  4. import random
  5. import numpy as np
  6. import onnxruntime as ort
  7. import onnx
  8. import torch
  9.  
  10. # import torchvision
  11. # from PIL import Image
  12. # from pathlib import Path
  13. # from collections import OrderedDict, namedtuple
  14.  
  15. cuda = torch.cuda.is_available()
  16. w = "bestE2E.onnx"  # 模型名称
  17.  
  18. # 判断模型是否有效，这步可以省略
  19. onnx_model = onnx.load(w)
  20. try:
  21.     onnx.checker.check_model(onnx_model)
  22. except Exception:
  23.     print("incorrect model")
  24.     exit()
  25. else:
  26.     print("correct model")
  27.  
  28. # 创建session
  29. providers = ['CUDAExecutionProvider', 'CPUExecutionProvider'] if cuda else ['CPUExecutionProvider']
  30. session = ort.InferenceSession(w, providers=providers)
  31.  
  32.  
  33. # Resize原始图像，预处理
  34. def letterbox(im, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleup=True, stride=32):
  35.     # Resize and pad image while meeting stride-multiple constraints
  36.     shape = im.shape[:2]  # current shape [height, width]
  37.     if isinstance(new_shape, int):
  38.         new_shape = (new_shape, new_shape)
  39.  
  40.     # Scale ratio (new / old)
  41.     r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
  42.     if not scaleup:  # only scale down, do not scale up (for better val mAP)
  43.         r = min(r, 1.0)
  44.  
  45.     # Compute padding
  46.     new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
  47.     dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
  48.  
  49.     if auto:  # minimum rectangle
  50.         dw, dh = np.mod(dw, stride), np.mod(dh, stride)  # wh padding
  51.  
  52.     dw /= 2  # divide padding into 2 sides
  53.     dh /= 2
  54.     print(shape[::-1])
  55.     if shape[::-1] != new_unpad:  # resize
  56.         im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)
  57.     top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
  58.     left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
  59.     im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)  # add border
  60.     return im, r, (dw, dh)
  61.  
  62.  
  63. # 类别名称，根据需要自己修改
  64. names = ['bubbles']
  65. colors = {name: [random.randint(0, 255) for _ in range(3)] for i, name in enumerate(names)}
  66.  
  67.  
  68. # 读入待检测图像
  69. img_path = r"D:/YOLOV7/img/test.bmp"
  70. img = cv2.imread(img_path)
  71.  
  72. # BGR to RGB，并增加一维
  73. image = img.copy()
  74. image, ratio, dwdh = letterbox(image, auto=False)
  75. image = image.transpose((2, 0, 1))
  76. image = np.expand_dims(image, 0)
  77. image = np.ascontiguousarray(image)
  78.  
  79. # 归一化
  80. im = image.astype(np.float32)
  81. im /= 255
  82.  
  83. outname = [i.name for i in session.get_outputs()]
  84. inname = [i.name for i in session.get_inputs()]
  85.  
  86. # 调用模型得出结果，并记录时间
  87. inp = {inname[0]: im}
  88. t1 = time.time()
  89. outputs = session.run(outname, inp)[0]
  90. print('inference time :%.4f' % (time.time() - t1))
  91.  
  92. ori_images = [img.copy()]
  93.  
  94. # 可视化结果
  95. for i, (batch_id, x0, y0, x1, y1, cls_id, score) in enumerate(outputs):
  96.     image = ori_images[int(batch_id)]
  97.     box = np.array([x0, y0, x1, y1])
  98.     box -= np.array(dwdh * 2)
  99.     box /= ratio
  100.     box = box.round().astype(np.int32).tolist()
  101.     cls_id = int(cls_id)
  102.     score = round(float(score), 3)
  103.     name = names[cls_id]
  104.     color = colors[name]
  105.     name += ' ' + str(score)
  106.     cv2.rectangle(image, box[:2], box[2:], color, 2)
  107.     cv2.putText(image, name, (box[0], box[1] - 2), cv2.FONT_HERSHEY_SIMPLEX, 0.75, [225, 255, 255], thickness=2)
  108.  
  109. cv2.imshow('dddd', ori_images[0])
  110. cv2.waitKey(0)
  111. cv2.destroyAllWindows()
  112.

结果显示：

二，C# 版

 公主做的demo界面如下：

 步骤与python版一模一样，公主一句一句翻译过来的。中间遇到过很多坑，但事情太多忘记一条一条记录了，有问题随时欢迎评论区讨论，完整代码如下：

  1. private void btn_analyze_Click(object sender, EventArgs e){
  2.             
  3.             // 原始图像切割为3份，分成1024*1024
  4.             Mat img_img = this.img_mat;
  5.             Mat[] subImg = new Mat[3];
  6.             int w = Width / 3;
  7.  
  8.             subImg[0] = img_img[0, 1024, startX, startX + w];
  9.             subImg[1] = img_img[0, 1024, startX + w, startX + w * 2];
  10.             subImg[2] = img_img[0, 1024, startX + w * 2, startX + rect.Width - edge * 2];
  11.  
  12.             // pictureBox对象“们”存为数组
  13.             PictureBox[] picbox = new PictureBox[3] { this.pictureBox2, this.pictureBox3, this.pictureBox4 };
  14.             for(int i = 0; i < 3; i++)
  15.             {
  16.                 int top = 0, left = 0;
  17.                 double ratio = 0.0;
  18.                 // after resize and padding
  19.                 Mat img_copy = new Mat();
  20.                 subImg[i].CopyTo(img_copy);
  21.                 
  22.                 // letterbox实现Resize图像
  23.                 Mat imageRgb = letterbox(img_copy, imgSize, out top, out left, out ratio);
  24.  
  25.                 // 将Mat格式图像转化为onnx模型需要的tensor张量
  26.                 Bitmap imageBmp = BitmapConverter.ToBitmap(imageRgb);
  27.                 Tensor<float> input = FastToOnnxTensor(imageBmp);
  28.                 var inputs = new List<NamedOnnxValue>
  29.                 {
  30.                     NamedOnnxValue.CreateFromTensor("images", input)
  31.                 };
  32.  
  33.                 // 调用onnx模型，保存结果
  34.                 IDisposableReadOnlyCollection<DisposableNamedOnnxValue> results = session.Run(inputs);
  35.  
  36.                 // 图像后处理，并可视化
  37.                 var resultsArray = results.ToArray()[0];
  38.                 Tensor<float> tensors = resultsArray.AsTensor<float>();
  39.                 var array = tensors.ToArray();
  40.  
  41.                 Mat newMat = new Mat();
  42.                 Cv2.CvtColor(subImg[i], newMat, ColorConversionCodes.GRAY2RGB);
  43.  
  44.                 if (array.Length > 0)
  45.                 {
  46.                     for (int j = 0; j < tensors.Dimensions[0]; j++)
  47.                     {
  48.                         int startIndex = j * tensors.Dimensions[1];
  49.                         Scalar color = new Scalar(127, 255, 0);  // 灰度颜色                        
  50.                         int lefttop_x = (int)((array[startIndex + 1] - left) / ratio);
  51.                         int lefttop_y = (int)((array[startIndex + 2] - top) / ratio);
  52.                         int rightbottom_x = (int)((array[startIndex + 3] - left) / ratio);
  53.                         int rightbottom_y = (int)((array[startIndex + 4] - top) / ratio);
  54.                         Cv2.Rectangle(newMat, new OpenCvSharp.Point(lefttop_x, lefttop_y),
  55.                                         new OpenCvSharp.Point(rightbottom_x, rightbottom_y),
  56.                                         color, 2);
  57.                         Cv2.PutText(newMat, LabelMap.Labels[(int)array[startIndex + 5]],
  58.                                     new OpenCvSharp.Point(lefttop_x, lefttop_y - 5),
  59.                                     HersheyFonts.HersheySimplex, 0.75, color, 2);
  60.                     }
  61.                 }
  62.                 picbox[i].Image = BitmapConverter.ToBitmap(newMat);
  63.             }
  64. }
  65.  
  66.  
  67. // Load表格时创建session
  68. private void OnnxDemo_Load(object sender, EventArgs e){
  69.             options = SessionOptions.MakeSessionOptionWithCudaProvider(0);
  70.  
  71.             session = new InferenceSession(path, options);
  72. }
  73.  
  74.  
  75. // letterbox预处理图像
  76. private Mat letterbox(Mat subImg, int newSize, out int top, out int left, out double ratio)
  77.         {
  78.             int width = subImg.Cols;
  79.             int height = subImg.Rows;
  80.  
  81.             ratio = Math.Min(((double)newSize / (double)width), ((double)newSize / (double)height));
  82.             int[] new_unpad = { (int)(Math.Round(height * ratio)), (int)(Math.Round(width * ratio)) };
  83.             int[] dwh = { (newSize - new_unpad[0]) / 2, (newSize - new_unpad[1]) / 2 };
  84.  
  85.             if (dwh != new_unpad)
  86.             {
  87.                 Cv2.Resize(subImg, subImg, new OpenCvSharp.Size(new_unpad[1], new_unpad[0]));
  88.             }
  89.  
  90.             top = (int)(Math.Round(dwh[0] - 0.1));
  91.             int bottom = newSize - top - new_unpad[0];
  92.             left = (int)(Math.Round(dwh[1] - 0.1));
  93.             int right = newSize - left - new_unpad[1];
  94.  
  95.             Cv2.CopyMakeBorder(subImg, subImg, top, bottom, left, right, BorderTypes.Constant, Scalar.Gray);
  96.  
  97.             Cv2.CvtColor(subImg, subImg, ColorConversionCodes.GRAY2RGB);  //Gray to RGB
  98.  
  99.             return subImg;
  100. }
  101.  
  102.  
  103. // Bitmap格式转化为Tensor张量
  104. public Tensor<float> FastToOnnxTensor(Bitmap source)
  105.         {
  106.             var floatArray = new float[source.Width * source.Height * 3];
  107.  
  108.             var bitmap_data = source.LockBits(new Rectangle(0, 0, source.Width, source.Height), ImageLockMode.ReadOnly, System.Drawing.Imaging.PixelFormat.Format24bppRgb);
  109.             var bitmap_bytes = new byte[Math.Abs(bitmap_data.Stride) * source.Height];
  110.  
  111.             Marshal.Copy(bitmap_data.Scan0, bitmap_bytes, 0, bitmap_bytes.Length);
  112.  
  113.             int total_pixels_count = source.Width * source.Height;
  114.  
  115.  
  116.             Parallel.For(0, total_pixels_count, (p_idx, state) =>
  117.             {
  118.  
  119.                 var g_idx = p_idx + total_pixels_count;
  120.                 var b_idx = p_idx + total_pixels_count * 2;
  121.  
  122.                 floatArray[p_idx] = bitmap_bytes[p_idx * 3] / 255f;//R
  123.                 floatArray[g_idx] = bitmap_bytes[p_idx * 3 + 1] / 255f;//G
  124.                 floatArray[b_idx] = bitmap_bytes[p_idx * 3 + 2] / 255f;//B
  125.  
  126.             });
  127.  
  128.             source.UnlockBits(bitmap_data);
  129.  
  130.             return new DenseTensor<float>(new Memory<float>(floatArray), new int[] { 1, 3, source.Height, source.Width });
  131.         }

静态变量：

  1. // 静态变量，用来保存类别名称
  2.  
  3. using System;
  4. using System.Collections.Generic;
  5. using System.Linq;
  6. using System.Text;
  7. using System.Threading.Tasks;
  8.  
  9. namespace OnnxDemo
  10. {
  11.     public static class LabelMap
  12.     {
  13.         static LabelMap()
  14.         {
  15.             Labels = new string[]
  16.             {
  17.                 "bubble"
  18.             };
  19.         }
  20.         public static string[] Labels { set; get; }
  21.     }
  22. }

运行结果：

 Nice~Nice~Nice~~~公主对结果还是灰常满意哒！！！

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/很楠不爱3/article/detail/295648?site
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号