Microsoft announced ML.NET 2.0, an updated version of its open-source machine learning framework ML.NET. A number of upgraded natural language processing (NLP) APIs, including tokenizers, text classification, and sentence similarity, are included in this release, along with enhanced automatic machine learning (AutoML) features.
The 2.0 version was announced by project manager Luis Quintanilla at the .NET Conf 2022. He briefed the attendees about the updated framework and added that it currently supports only NVIDIA’s CUDA GPU accelerators, but the developers will work to continue furthering the framework to increase compatibility with others in the future.
Currently, ML.NET 2.0 has been released with the following new features:
- The NLP APIs are powered by the latest .NET wrapper for PyTorch–TorchSharp.
- The new framework includes the EnglishRoberta model for tokenization.
- TorchSharp-driven implementation of NAS-BERT, wherein users can now fine-tune the NAS-BERT model using their own data and custom use cases. This model is a crucial element of both sentence similarity and text classification APIs.
- The framework also received updates to AutoML for data preprocessing and finding the most appropriate hyperparameters. These include Featurizer API, Experiment API, Sweepable API, Search Space API, and Tuner API.
Quintanilla also created a roadmap for the ML.NET framework and future scope of deep learning technologies for named-entity recognition, object detection, and question-answering. He also disclosed the addition of a new text categorization scenario and advanced training options in the latest version of the Model Builder tool for Visual Studio.
The ML.NET 2.0 framework also leaves room for more TorchSharp integrations and enhancements for ONYX-based integrations. Also, Microsoft plans to update the LightGBM implementation for IDataView interfaces.