This sample tells how to bind and evaluate batches of input in WinML

• Visual Studio 2017 - 15.4 or higher

• Windows 10 - Build 18362 or higher

• Windows SDK - Build 18362 or higher

• Visual Studio Extension for C++/WinRT

  Do the following to add the C++/WinRT extension in Visual Studio.

  1. Go to Tools > Extensions and Updates.
  2. Select Online in the left pane and search for "WinRT" using the search box.
  3. Select the C++/WinRT extension, click Download, and close Visual Studio. The extension should install automatically.
  4. When the extension has finished installing, re-open Visual Studio.

1. If you download the samples ZIP, be sure to unzip the entire archive, not just the folder with the sample you want to build.
2. Start Microsoft Visual Studio 2017 and select File > Open > Project/Solution.
3. Starting in the folder where you unzipped the samples, go to the Samples subfolder, then the subfolder for this specific sample (Sample\BatchSupport). Double-click the Visual Studio solution file (BatchSupport.sln).
4. Confirm that the project is pointed to the correct SDK that you installed (e.g. 18362). You can do this by right-clicking the project in the Solution Explorer, selecting Properties, and modifying the Windows SDK Version.
5. Confirm that you are set for the right configuration and platform (for example: Debug, x64).
6. Build the solution (Ctrl+Shift+B).

1. Open a Command Prompt (in the Windows 10 search bar, type cmd and press Enter).
2. Change the current folder to the folder containing the built EXE (cd <path-to-exe>).
3. Run the executable as shown below. Make sure to replace the install location with what matches yours:

BatchSupport.exe [fixedBatchSize|freeBatchSize] [TensorFloat|VideoFrame]

4. You should get output similar to the following:

   Loading modelfile 'E:\xianz\Windows-Machine-Learning\Samples\BatchSupport\x64\Debug\SqueezeNet.onnx' on the CPU
   model file loaded in 906 ticks
   Binding...
   Running the model...
   model run took 31 ticks
   Result for No.0 input
   tench, Tinca tinca with confidence of 0.738503
   Result for No.1 input
   tabby, tabby cat with confidence of 0.931461
   Result for No.2 input
   tench, Tinca tinca with confidence of 0.738503
   

1. Download WinMLDashboard
2. change the batch dimension of model input and output to a fixed number or -1 (free dimension)

Take binding batches of VideoFrame as example:

1.1 for fixed batch dimension:

 //create a session and binding
 LearningModelDeviceKind deviceKind = LearningModelDeviceKind::Cpu;
 LearningModelSession session(model, LearningModelDevice(deviceKind));
 LearningModelBinding binding(session);

1.2 for free-dimensional batch size:

 // create a session and binding
 LearningModelDeviceKind deviceKind = LearningModelDeviceKind::Cpu;
 LearningModelSessionOptions options;
 if ("freeBatchSize" == modelType) { 
     // If the model has free dimentional batch, override the free dimension with batch_size
     options.BatchSizeOverride(static_cast<uint32_t>(BATCH_SIZE));
 }
 LearningModelSession session(model, LearningModelDevice(deviceKind), options);
 LearningModelBinding binding(session);

    // Create Batches of VideoFrame
    std::vector<VideoFrame> inputVideoFrames = {};
    for (hstring imageName : imageNames) {
        auto imagePath = static_cast<hstring>(GetModulePath().c_str()) + imageName;
        auto imageFrame = LoadImageFile(imagePath);
        inputVideoFrames.emplace_back(imageFrame);
    }
    auto videoFrames = winrt::single_threaded_vector(std::move(inputFrames));

    // Bind
    binding.Bind(inputFeatureDescriptor.Current().Name(), inputVideoFrames);

The sample does not bind the output, but you could also bind the output as below:

  auto outputShape = std::vector<int64_t>{BATCH_SIZE, 1000, 1, 1};	
  auto outputValue = TensorFloat::Create(outputShape);	
  std::wstring outputDataBindingName =	
      std::wstring(model.OutputFeatures().First().Current().Name());	
  binding.Bind(outputDataBindingName, outputValue);
  SampleHelper::PrintResults(outputValue.GetAsVectorView()); // Print Results