The first preview release of ML.NET 3.0, available since December, contains the integration with Intel oneAPI Data Analytics Library that leverages SIMD extensions on 64-bit architectures, which are available on Intel and AMD processors.
ML.NET 3.0 allows importing more machine learning capabilities into existing .NET applications. This feature eliminates the need to code complex methods to generate predictive models.
Intel oneAPI Data Analytics Library helps accelerate data analysis of large datasets by providing highly optimised algorithmic building blocks for all stages of the data analysis such as preprocessing, modelling or decision-making. These blocks are independent of any communication layer. The library optimises data acquisition along with algorithmic calculations to increase scalability and throughput.
Jake Radzikowski and Luis Quintanilla mentioned in the last Machine Learning Community Standup that the ML.NET team worked closely with Intel to optimise Intel hardware, but it is also compatible with AMD CPUs. ML.NET 3.0 brings some new algorithms and improves the existing ones. The next step of the ML.NET 3.0 development will be integrating these algorithms with AutoML and Model Builder. More details about the upcoming features and improvements are accessible on the ML.NET roadmap.
oneAPI Data Analytics Library integrates with ML.NET, speeding up the training of the following trainers: L-BGFS, Ordinary Least Squares, FastTree and FastForest. L-BGFS is used for regression tasks, Ordinary Least Squares for classification and the last two for both regression and classification.
In order to start using oneAPI Data Analytics Library in ML.NET, it is necessary to install the latest Microsoft.ML 3.0 preview version and the Microsoft.ML.OneDal NuGet package. In addition, a key element is setting the MLNET_BACKEND environment variable to ONEDAL. There is also a known problem on Windows with the library loading. In order to fix this, it is necessary to add the 'runtimes-x64native' directory in the 'bin' directory of the application to the PATH environment variable.
A usage example is available on GitHub.
