TMVA SOFIE: Enhancing the Machine Learning Inference Engine

ROOT [4] is primarily used for high-scale data processing and analysis required for the research of high-energy physics. It provides various tools, utilities, and facilities for statistics, algebra, visualization, multi-varied methods, data serialization, parallelized data computation, etc. With the advent of machine learning, it is nowadays necessary to use regression and classification methods for the determination and analysis of physical events. Consequently, ROOT offers native support for various supervised learning techniques and easy interoperability with commonly used machine learning libraries through its module called the Toolkit of Multi-varied Analysis, or TMVA [5] . In the recent developments in TMVA, research on machine learning inference was being made for developing a Fast Machine Learning Inference Engine called SOFIE [1] . SOFIE which stands for System for Optimized Fast Inference code Emit was launched in 2021, and since then active work has been going on to improve its capabilities. SOFIE is based on ONNX [7] standards and presents a system that generates inference code with the least latency and few dependencies, suitable for applications in high-energy physics. SOFIE has support for parsing models developed in ONNX, Keras, or PyTorch framework. On this project of enhancing the inference engine, rigorous work was done on strengthening the Keras parser, designing and developing the support for a custom operator, and implementing the functionality for parsing and inference of graph neural networks in SOFIE.