An update to the machine learning framework for .NET developers brings new capabilities in object detection, named entity recognition, and question answering.
Microsoft has released ML.NET 3.0, the latest version of the company’s open-source, cross-platform machine learning framework, which enables the integration of machine learning models into .NET applications.
Announced November 27, ML.NET 3.0 can be accessed from dotnet.microsoft.com.
ML.NET 3.0 includes new deep learning capabilities in object detection, named entity recognition, and question answering. These deep-learning scenarios were supported via integrations and interoperability with TorchSharp and ONNX models. The 3.0 release also updates integration with the LightGBM gradient boosting framework.
ML.NET 3.0 also improves support for data processing scenarios with enhancements and bug fixes to DataFrame and new IDataView interoperability features. Loading, inspecting, transforming, and visualizing data have been made more powerful.
In May, Microsoft announced object detection in ML.NET Model Builder. These capabilities are built on top of TorchSharp-powered Object Detection APIs introduced in ML.NET 3.0. The Object Detection API leverages some of the latest techniques from Microsoft Research and is backed by a Transformer-based neural network architecture built with TorchSharp. Object detection is included in the Microsoft ML.TorchSharp 3.0.0 package.
ML.NET 3.0 also offers natural language processing areas, including question answering and named entity recognition. These scenarios are unlocked by building on top of the existing TorchSharp RoBERTa text classification features introduced in ML.NET 2.0. And ML.NET 3.0 gains new automated machine learning (AutoML) capabilities including the AutoML Sweeper now supporting sentence similarity, question answering, and object detection.
DataFrame updates have been made in ML.NET 3.0, including expanded data loading scenarios, with data now importable from and exportable to SQL databases. This is done via ADO.NET, which supports SQL-compatible databases. Also in DataFrame, arithmetic performance was improved in column cloning and binary comparison scenarios. Null value handling was improved while performing arithmetic operations, requiring fewer steps in transforming data. Debugger improvements were made to improve readable output for columns with long names. Tensor Primitives include a new set of APIs to support tensor operations.
Microsoft now is working on plans for .NET 9 and ML.NET 4.0. In the meantime, the company said users can expect Model Builder and the ML.NET CLI to be updated to consume the ML.NET 3.0 release. Plans also call for expanding deep learning scenarios and integrations and enhancing DataFrame. Finally, Microsoft said it would continue to expand the APIs in System.Numerics.Tensors and integrate them into ML.NET.