What it does This repository holds sample applications and developer tools for the original Windows Machine Learning API (Windows.AI.MachineLearning), shipped in 2018 as part of the Windows SDK. The samples demonstrate how to run ONNX model inference in UWP and desktop Windows apps, covering image classification with models like SqueezeNet and ResNet50, style transfer, and advanced scenarios such as custom tensorization and adapter selection.

The interesting bit The repo is essentially a museum of Windows ML deployment options. It catalogs the same models across four permutations: UWP vs. desktop, inbox SDK vs. NuGet package. The most unexpected entry is a Rust sample for SqueezeNet object detection—an outlier in an otherwise C++/C#/JavaScript collection.

Key highlights

• Built on ONNX Runtime and DirectML for GPU-accelerated inference
• Supports both inbox Windows SDK (Windows 10 RS5+) and redistributable NuGet (back to Windows 8.1)
• Includes tools: WinmlRunner for command-line model testing and WinML Dashboard for model management
• Advanced samples cover custom CPU operators, adapter selection, and custom tensorization
• Multiple language bindings: C++/WinRT, C++/CX, C#, JavaScript, and Rust

Caveats

• The README prominently warns that these APIs are superseded by Microsoft.Windows.AI.MachineLearning in the Windows App SDK; this repo documents the old namespace
• Visual Studio 2017 15.7.4 or newer is listed as a prerequisite, suggesting the build infrastructure hasn’t been refreshed for current toolchains
• Some APIs like VideoFrame are unavailable on older OSes even with the NuGet package

Verdict Worth browsing if you’re maintaining legacy Windows 10 RS5-era applications or need to compare the old and new Windows ML API shapes. Skip it if you’re starting fresh—Microsoft has already pointed the way to the Windows App SDK successor.