Windows-Machine-Learning - High-Performance Machine Learning Inference Powered by ONNX Runtime and DirectML

Windows Machine Learning: A Comprehensive Introduction

Overview

Windows Machine Learning (Windows ML) is a high-performance API designed to facilitate machine learning inference. It harnesses the power of ONNX Runtime and DirectML to deliver enhanced performance and low latency suitable for a variety of applications, including frameworks, games, and real-time applications. This robust API is well-equipped to integrate into applications using high-level languages.

Getting Started

Prerequisites:

To get started with Windows ML, developers need to have Visual Studio 2017 Version 15.7.4 or newer installed. The Windows ML inference is available via the inbox Windows SDK and as a redistributable NuGet package. It's crucial to understand the differences in availability, distribution, compatibility, and servicing between the In-Box and NuGet options. For instance, the In-Box version is built into Windows and automatically updates with new features, while the NuGet version requires manual updates by developers.

Model Samples

Windows ML offers various model samples demonstrating its capabilities:

Image Classification: This involves categorizing images into predefined tags. Models such as AlexNet, DenseNet, and SqueezeNet are among the popular ones supported, showcasing their use in different Windows application types, including UWP and Desktop applications.
Style Transfer: This fascinating technique reimagines an image's content using the style of another image. Examples include the FNSCandy and Real-Time Style Transfer apps developed in C#.

Advanced Scenario Samples

For those looking to explore deeper functionalities, Windows ML provides samples for advanced scenarios such as:

Custom Tensorization: Showcases custom tensorization in a Windows Console Application.
Custom Operator (CPU): Enables the definition of custom CPU operators.
Adapter Selection: Demonstrates device adapter selection for model execution.
Integration with Cloud Services: Tutorials on deploying models trained using Azure Custom Vision or ML.NET in Windows ML applications.
PyTorch Tutorials: Guides on training models using PyTorch, converting to ONNX, and deploying as Windows ML applications.

Developer Tools

Windows ML extends support with developer tools for model enhancement and deployment:

Model Conversion and Optimization: Windows ML relies on the ONNX model format. Tools like WinMLTools facilitate conversion from frameworks like TensorFlow or PyTorch. Further optimizations, such as batching and quantization, can be applied post-conversion via tools like WinML Dashboard and ONNX Runtime Graph Optimizations.
Model Validation and Integration: WinMLRunner aids in validating ONNX models, while the WinML Code Generator helps integrate models into UWP applications. Developers can further explore diverse ML integration scenarios through the WinML Samples Gallery.

Feedback and Contribution

The Windows ML project encourages engagement from its community, welcoming feedback through GitHub Issues and participation in questions on Stack Overflow. Contributions are also invited, allowing developers to help improve samples and fix bugs.

Conclusion

Windows Machine Learning is a versatile and powerful toolset for developers interested in machine learning applications on Windows platforms. By providing a comprehensive suite of tools, samples, and advanced features, Windows ML supports a wide range of development activities, from beginners to seasoned ML practitioners. For more information, the Windows ML documentation is a valuable resource.