API¶
graphnet: Graph neural networks for neutrino telescope event reconstruction.
graphnet is a python package that provides convenient, common, and collaboratively developed tools for building graph neural networks (GNNs) to solve physics tasks at neutrino telescope experiments. It aims to provide physicists with the tools to leverage advanced machine learning (ML) without having to be machine learning experts themselves, and thereby accelerate the scientific advances in the area of neutrino phyics.
Design principles:
End-to-end: graphnet aims to provide all of the tools for streamlining the process of ingesting and transforming physics data; building, training, and optimising GNN models; and deploying them into a reconstruction chain.
Extensibility: graphnet aims to provide the basic building blocks to improve reconstruction and classification across the various IceCube configurations, but all model components can be easily extended to new experiments, with new GNN architectures and for new physics tasks.
Main features:
Converters from domain-specific data formats (I3) to more common, indexable formats (e.g., SQLite) suitable as intermediate file formats for training ML models
Plug-and-play GNN model components that abstract away ML implementation details and only expose the “building blocks” that are most relevant to physicsts (e.g., what detector is used and what are the physicst tasks).
I3Modules for easily including GNN models in IceCube reconstruction chains.
Docker images for running model inference in a containerised fashion.