使用 Windows 机器学习 API 在 Windows 应用中部署 TensorFlow 模型

项目
07/11/2023

最后一部分将解释如何创建带有 GUI 的简单 UWP 应用，通过使用 Windows ML 评估 YOLO 模型来流式传输网络摄像头和检测物体。

在 Visual Studio 中创建 UWP 应用

打开 Visual Studio，选择 Create a new project. 搜索 UWP，然后选择 Blank App (Universal Windows)。

Create a new project for your app

在下一页上，通过为项目指定名称和位置来配置项目设置。然后选择应用的目标和最低 OS 版本。要使用 Windows ML API，则必须使用 X，也可选择 NuGet 包以向下支持 X。如果选择使用 NuGet 包，请按照以下说明操作：[link]。

调用 Windows ML API 来评估模型

步骤 1：使用机器学习代码生成器为 Windows ML API 生成包装器类。

步骤 2：在生成的 .cs 文件中修改生成的代码。最终文件如下所示：

using System;
using System.Collections.Generic;
using System.Threading.Tasks;
using Windows.Media;
using Windows.Storage;
using Windows.Storage.Streams;
using Windows.AI.MachineLearning;
namespace yolodemo
{
    
    public sealed class YoloInput
    {
        public TensorFloat input_100; // shape(-1,3,416,416)
    }
    
    public sealed class YoloOutput
    {
        public TensorFloat concat_1600; // shape(-1,-1,-1)
    }
    
    public sealed class YoloModel
    {
        private LearningModel model;
        private LearningModelSession session;
        private LearningModelBinding binding;
        public static async Task<YoloModel> CreateFromStreamAsync(IRandomAccessStreamReference stream)
        {
            YoloModel learningModel = new YoloModel();
            learningModel.model = await LearningModel.LoadFromStreamAsync(stream);
            learningModel.session = new LearningModelSession(learningModel.model);
            learningModel.binding = new LearningModelBinding(learningModel.session);
            return learningModel;
        }
        public async Task<YoloOutput> EvaluateAsync(YoloInput input)
        {
            binding.Bind("input_1:0", input.input_100);
            var result = await session.EvaluateAsync(binding, "0");
            var output = new YoloOutput();
            output.concat_1600 = result.Outputs["concat_16:0"] as TensorFloat;
            return output;
        }
    }
}

评估每个视频帧以检测物体并绘制范围框。

将以下库添加到 mainPage.xaml.cs。

using System.Threading.Tasks;
using Windows.Devices.Enumeration;
using Windows.Media;
using Windows.Media.Capture;
using Windows.Storage;
using Windows.UI;
using Windows.UI.Xaml.Media.Imaging;
using Windows.UI.Xaml.Shapes;
using Windows.AI.MachineLearning;

在 public sealed partial class MainPage : Page 中添加以下变量。

private MediaCapture _media_capture;
private LearningModel _model;
private LearningModelSession _session;
private LearningModelBinding _binding;
private readonly SolidColorBrush _fill_brush = new SolidColorBrush(Colors.Transparent);
private readonly SolidColorBrush _line_brush = new SolidColorBrush(Colors.DarkGreen);
private readonly double _line_thickness = 2.0;
        private readonly string[] _labels =
            {
                "<list of labels>"
        };

创建表现检测结果格式的结构。

        internal struct DetectionResult
        {
            public string label;
            public List<float> bbox;
            public double prob;
        }

创建一个 Comparer 对象，用于比较两个 Box 类型的对象。此类将用于在检测到的物体周围绘制范围框。

        class Comparer : IComparer<DetectionResult>
        {
            public int Compare(DetectionResult x, DetectionResult y)
            {
                return y.prob.CompareTo(x.prob);
            }
        }

添加以下方法以初始化设备的网络摄像头流，并开始处理每一帧以检测物体。

        private async Task InitCameraAsync()
        {
            if (_media_capture == null || _media_capture.CameraStreamState == Windows.Media.Devices.CameraStreamState.Shutdown || _media_capture.CameraStreamState == Windows.Media.Devices.CameraStreamState.NotStreaming)
            {
                if (_media_capture != null)
                {
                    _media_capture.Dispose();
                }

                MediaCaptureInitializationSettings settings = new MediaCaptureInitializationSettings();
                var cameras = await DeviceInformation.FindAllAsync(DeviceClass.VideoCapture);
                var camera = cameras.FirstOrDefault();
                settings.VideoDeviceId = camera.Id;

                _media_capture = new MediaCapture();
                await _media_capture.InitializeAsync(settings);
                WebCam.Source = _media_capture;
            }

            if (_media_capture.CameraStreamState == Windows.Media.Devices.CameraStreamState.NotStreaming)
            {
                await _media_capture.StartPreviewAsync();
                WebCam.Visibility = Visibility.Visible;
            }

            ProcessFrame();
        }

添加以下方法来处理每一帧。此方法调用 EvaluateFrame 和 DrawBoxes，我们将在后面的步骤中实现它们。

        private async Task ProcessFrame()
        {
            var frame = new VideoFrame(Windows.Graphics.Imaging.BitmapPixelFormat.Bgra8, (int)WebCam.Width, (int)WebCam.Height);
            await _media_capture.GetPreviewFrameAsync(frame);
            var results = await EvaluateFrame(frame);
            await DrawBoxes(results.ToArray(), frame);
            ProcessFrame();
        }

创建新的 Sigmoid 浮点

        private float Sigmoid(float val)
        {
            var x = (float)Math.Exp(val);
            return x / (1.0f + x);
        }

创建用于正确检测物体的阈值。

        private float ComputeIOU(DetectionResult DRa, DetectionResult DRb)
        {
            float ay1 = DRa.bbox[0];
            float ax1 = DRa.bbox[1];
            float ay2 = DRa.bbox[2];
            float ax2 = DRa.bbox[3];
            float by1 = DRb.bbox[0];
            float bx1 = DRb.bbox[1];
            float by2 = DRb.bbox[2];
            float bx2 = DRb.bbox[3];

            Debug.Assert(ay1 < ay2);
            Debug.Assert(ax1 < ax2);
            Debug.Assert(by1 < by2);
            Debug.Assert(bx1 < bx2);
   
            // determine the coordinates of the intersection rectangle
            float x_left = Math.Max(ax1, bx1);
            float y_top = Math.Max(ay1, by1);
            float x_right = Math.Min(ax2, bx2);
            float y_bottom = Math.Min(ay2, by2);

            if (x_right < x_left || y_bottom < y_top)
                return 0;
            float intersection_area = (x_right - x_left) * (y_bottom - y_top);
            float bb1_area = (ax2 - ax1) * (ay2 - ay1);
            float bb2_area = (bx2 - bx1) * (by2 - by1);
            float iou = intersection_area / (bb1_area + bb2_area - intersection_area);

            Debug.Assert(iou >= 0 && iou <= 1);
            return iou;
        }

实现以下列表，以跟踪帧中检测到的当前物体。

        private List<DetectionResult> NMS(IReadOnlyList<DetectionResult> detections, 
            float IOU_threshold = 0.45f, 
            float score_threshold=0.3f)
        {
            List<DetectionResult> final_detections = new List<DetectionResult>();
            for (int i = 0; i < detections.Count; i++)
            {
                int j = 0;
                for (j = 0; j < final_detections.Count; j++)
                {
                    if (ComputeIOU(final_detections[j], detections[i]) > IOU_threshold)
                    {
                        break;
                    }
                }
                if (j==final_detections.Count)
                {
                    final_detections.Add(detections[i]);
                }
            }
            return final_detections;
        }

实现以下方法。

        private List<DetectionResult> ParseResult(float[] results)
        {
            int c_values = 84;
            int c_boxes = results.Length / c_values;
            float confidence_threshold = 0.5f;
            List<DetectionResult> detections = new List<DetectionResult>();
            this.OverlayCanvas.Children.Clear();
            for (int i_box = 0; i_box < c_boxes; i_box++)
            {
                float max_prob = 0.0f;
                int label_index = -1;
                for (int j_confidence = 4; j_confidence < c_values; j_confidence++)
                {
                    int index = i_box * c_values + j_confidence;
                    if (results[index] > max_prob)
                    {
                        max_prob = results[index];
                        label_index = j_confidence - 4;
                    }
                }
                if (max_prob > confidence_threshold)
                {
                    List<float> bbox = new List<float>();
                    bbox.Add(results[i_box * c_values + 0]);
                    bbox.Add(results[i_box * c_values + 1]);
                    bbox.Add(results[i_box * c_values + 2]);
                    bbox.Add(results[i_box * c_values + 3]);

                    detections.Add(new DetectionResult()
                    {
                        label = _labels[label_index],
                        bbox = bbox,
                        prob = max_prob
                    });
                }
            }
            return detections;
        }

添加以下方法，以在帧中检测到的物体周围绘制框。

private async Task DrawBoxes(float[] results, VideoFrame frame)
        {
            List<DetectionResult> detections = ParseResult(results);
            Comparer cp = new Comparer();
            detections.Sort(cp);
            IReadOnlyList<DetectionResult> final_detetions = NMS(detections);

            for (int i=0; i < final_detetions.Count; ++i)
            {
                int top = (int)(final_detetions[i].bbox[0] * WebCam.Height);
                int left = (int)(final_detetions[i].bbox[1] * WebCam.Width);
                int bottom = (int)(final_detetions[i].bbox[2] * WebCam.Height);
                int right = (int)(final_detetions[i].bbox[3] * WebCam.Width);

                var brush = new ImageBrush();
                var bitmap_source = new SoftwareBitmapSource();
                await bitmap_source.SetBitmapAsync(frame.SoftwareBitmap);

                brush.ImageSource = bitmap_source;
                // brush.Stretch = Stretch.Fill;

                this.OverlayCanvas.Background = brush;

                var r = new Rectangle();
                r.Tag = i;
                r.Width = right - left;
                r.Height = bottom - top;
                r.Fill = this._fill_brush;
                r.Stroke = this._line_brush;
                r.StrokeThickness = this._line_thickness;
                r.Margin = new Thickness(left, top, 0, 0);

                this.OverlayCanvas.Children.Add(r);
                // Default configuration for border
                // Render text label
                

                var border = new Border();
                var backgroundColorBrush = new SolidColorBrush(Colors.Black);
                var foregroundColorBrush = new SolidColorBrush(Colors.SpringGreen);
                var textBlock = new TextBlock();
                textBlock.Foreground = foregroundColorBrush;
                textBlock.FontSize = 18;

                textBlock.Text = final_detetions[i].label;
                // Hide
                textBlock.Visibility = Visibility.Collapsed;
                border.Background = backgroundColorBrush;
                border.Child = textBlock;

                Canvas.SetLeft(border, final_detetions[i].bbox[1] * 416 + 2);
                Canvas.SetTop(border, final_detetions[i].bbox[0] * 416 + 2);
                textBlock.Visibility = Visibility.Visible;
                // Add to canvas
                this.OverlayCanvas.Children.Add(border);
            }
        }

现在我们处理了必要的基础结构，接下来合并评估结果。此方法根据当前帧评估模型，以检测物体。

        private async Task<List<float>> EvaluateFrame(VideoFrame frame)
        {
            _binding.Clear();
            _binding.Bind("input_1:0", frame);
            var results = await _session.EvaluateAsync(_binding, "");
            Debug.Print("output done\n");

            TensorFloat result = results.Outputs["Identity:0"] as TensorFloat;
            var shape = result.Shape;
            var data = result.GetAsVectorView();
            
            return data.ToList<float>();
        }

应用需要以某种方式启动。添加一个方法，它在用户按下 Go 按钮时开始网络摄像头流和模型评估。

        private void button_go_Click(object sender, RoutedEventArgs e)
        {
            InitModelAsync();
            InitCameraAsync();
        }

添加一个方法，用于调用 Windows ML API 来评估模型。首先从存储中加载模型，然后创建会话并将其绑定到内存。

        private async Task InitModelAsync()
        {
            var model_file = await StorageFile.GetFileFromApplicationUriAsync(new Uri("ms-appx:///Assets//Yolo.onnx"));
            _model = await LearningModel.LoadFromStorageFileAsync(model_file);
            var device = new LearningModelDevice(LearningModelDeviceKind.Cpu);
            _session = new LearningModelSession(_model, device);
            _binding = new LearningModelBinding(_session);
        }

启动应用程序

你已成功创建实时物体检测应用程序！选择 Visual Studio 顶部栏上的 Run 按钮来启动应用。该应用应如下所示。

The final TensorFlow app

其他资源

若要详细了解本教程中所述的主题，请访问以下资源:

Windows ML 工具：了解更多工具，例如 Windows ML 仪表板、WinMLRunner 和 mglen Windows ML 代码生成器。
ONNX 模型：详细了解 ONNX 格式。
Windows ML 性能和内存：详细了解如何使用 Windows ML 管理应用性能。
Windows 机器学习 API 参考：详细了解 Windows ML API 的三个方面。

通过

使用 Windows 机器学习 API 在 Windows 应用中部署 TensorFlow 模型

在 Visual Studio 中创建 UWP 应用

调用 Windows ML API 来评估模型

评估每个视频帧以检测物体并绘制范围框。

启动应用程序

其他资源

反馈

其他资源