🎯 How to call Microsoft.Windows.AI from C++ [With Sample Code]

Since Windows 10, Windows has been equipped with a built-in **runtime capable of executing ONNX format AI models . This is ** Windows ML (Windows.AI.MachineLearning).

In this article, we will specifically explain how to call Microsoft.Windows.AI.MachineLearning from C++ (Win32 app based), along with ** sample code**.

✅ Preparation

◾ System Requirements

• Windows 10 (1809+) or Windows 11
• Visual Studio 2019 or later (Community edition is fine)
• C++/WinRT Support (Microsoft.Windows.CppWinRT)
• Windows SDK 10.0.17763.0 or higher

✅ Project Configuration

Create a project in Visual Studio with the following configuration:

• Type: C++ Windows Desktop Application (Empty Project)

• Subsystem: Windows (WinMain)

• Add the following package via NuGet:

  1	Microsoft.Windows.CppWinRT

✅ Sample Code

Below is a minimal sample combining the Win32 API and Windows.AI.MachineLearning using WinMain.

• Note: Assume the ONNX model to be used is model.onnx, and place it in the same folder as the executable file.

main.cpp

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38

#include <windows.h>
#include <winrt/Windows.AI.MachineLearning.h>
#include <winrt/Windows.Storage.h>

#pragma comment(lib, "windowsapp") // For WinRT linking

using namespace winrt;
using namespace Windows::AI::MachineLearning;
using namespace Windows::Storage;

int WINAPI WinMain(HINSTANCE hInstance, HINSTANCE, LPSTR, int nCmdShow)
{
    // Initialize WinRT (Either MTA or STA is fine)
    winrt::init_apartment();

    try {
        // Load the model file
        auto modelFile = StorageFile::GetFileFromPathAsync(L"model.onnx").get();
        LearningModel model = LearningModel::LoadFromStorageFileAsync(modelFile).get();

        // Create a session
        LearningModelSession session(model);
        LearningModelBinding binding(session);

        // Model input/output (Here, a temporary empty input)
        // In practice, binding with TensorFloat etc. is required

        // Execute inference
        auto result = session.EvaluateAsync(binding, L"").get();

        MessageBox(nullptr, L"Inference completed", L"Windows ML (C++)", MB_OK);
    }
    catch (winrt::hresult_error const& ex) {
        MessageBox(nullptr, ex.message().c_str(), L"Error", MB_ICONERROR);
    }

    return 0;
}

✅ Supplement: How to Specify Input/Output Tensors

Depending on the model, it is necessary to create and bind Tensors before inference.

Example:

1
2
3
4
5
6
7
8

// Convert a 1D float array to a Tensor
std::vector<float> inputData = {0.5f, 0.3f, 0.2f};
std::vector<int64_t> shape = {1, 3}; // Shape: [1, 3]

auto tensor = TensorFloat::CreateFromArray(shape, inputData);

// Input binding (Match the model's input name)
binding.Bind(L"input_0", tensor);

Outputs can be obtained similarly using result.Outputs().Lookup(L"output_0").

✅ Debugging Tips

• A FileNotFoundException will be thrown if the model file is not in the execution folder.
• An invalid_argument error will occur if the input/output names do not match.
• The exact I/O specifications of the model can be confirmed with tools like Netron.

✅ Summary

✅ Alternatives: For those who do not want to use WinRT

• By using Microsoft’s ONNX Runtime, you can handle ONNX models from C++ entirely without WinRT.
• It supports cross-platform, allowing common code for Windows/Linux.

📌 Conclusion

Windows ML (Microsoft.Windows.AI) is a powerful AI inference engine that can be robustly used even from C++. If you need native inference on Windows, please give it a try.

For those who want specific examples of creating ONNX models and Tensor binding, we plan to explain them in a follow-up article!