Exa.TrkX plugin
This plugin contains a track finding module based on Graph Neural Networks (GNNs) which is developed by the Exa.TrkX team. Build instructions and dependencies can be found in the README of the plugin.
Stages
The Exa.TrkX pipeline is a multi-stage GNN-based algorithm. In principle, there are three types of stages:
Graph construction: This stage constructs the initial graph from the space points. Currently, there is only a metric-learning based approach implemented: A neural network tries to learn a mapping that minimizes the distance between points of the same track in the embedded space. In this embedded space then a graph is built using a fixed nearest-neighbor search.
Edge classification: In this stage, a graph is taken from the previous stage, and an edge-classification is performed on the edges. This can be done either by a simple feed forward network or by a GNN.
Track building stage: In this stage, track candidates are built from the edges and the scores of the previous edge classification stage. Currently, there are simple track building algorithms built on top of a weakly connected components algorithm available.
A typical pipeline consists e.g. of 4 stages:
Graph construction: Metric learning
|
v
Edge classification: Filter
|
V
Edge classification: GNN
|
V
Track building with boost::graph
Implementation
Note
The codebase is currently under refactoring, the documentation will be updated once the code has stabilized.
See here for the corresponding python example.
Resources
Talk by Daniel Murnane at the Connecting the Dots 2020
Talk by Daniel Murnane at the vCHEP 2021
Talk by Alina Lazar at the ACAT 2021
Talk by Benjamin Huth at the ICHEP 2022