Windows.ai.machinelearning 🎯 Ad-Free

var session = new LearningModelSession(model, device);

// Prepare input tensor (example: image 224x224 RGB) var inputData = new float[1 * 3 * 224 * 224]; // fill with your image data var inputTensor = TensorFloat.CreateFromArray(new long[] 1, 3, 224, 224 , inputData); binding.Bind("input", inputTensor);

using Microsoft.ML.OnnxRuntime; using Microsoft.AI.MachineLearning; // Load model var file = await StorageFile.GetFileFromApplicationUriAsync( new Uri("ms-appx:///Assets/model.onnx")); var model = await LearningModel.LoadFromStorageFileAsync(file); // Create session var session = new LearningModelSession(model, new LearningModelDevice(LearningModelDeviceKind.Default)); // Create binding var binding = new LearningModelBinding(session); windows.ai.machinelearning

// 1. Preprocess: resize to model input size (224x224) var resized = await ImageHelper.ResizeBitmap(bitmap, 224, 224); // 2. Convert to float tensor (channel-first, normalized) var tensor = ImageHelper.BitmapToTensor(resized);

// Run inference var results = await session.EvaluateAsync(binding, "runId"); Load model (cache globally) var model = await App

// 3. Load model (cache globally) var model = await App.ModelLoader.GetModelAsync();

// Force GPU var device = new LearningModelDevice(LearningModelDeviceKind.DirectXHighPerformance); // Force NPU (Windows 11 24H2+) var device = new LearningModelDevice(LearningModelDeviceKind.Npu); options.CloseModelOnSessionCreation = false

// 4. Bind & evaluate var session = new LearningModelSession(model); var binding = new LearningModelBinding(session); binding.Bind("data", tensor);

// 5. Map to label return Labels[ArgMax(classId)]; Windows ML automatically uses DirectML – you don’t need to change code. But you can select the device:

LearningModelSessionOptions options = new LearningModelSessionOptions(); options.CloseModelOnSessionCreation = false; options.LoggingName = "MyModel";

var result = await session.EvaluateAsync(binding, ""); var classId = result.Outputs["softmaxout"] as TensorFloat;