Speaker
Xiangyang Ju
(Lawrence Berkeley National Lab. (US))
Description
We developed a python-based package that facilitates the usage of graph neural network on HEP data. It is featured with pre-defined GNN models for edge
classification and event classification. It also
contains a couple of realistic examples using GNN to solve HEP
problems, for example, top tagger (event classification) and boosted
boson reconstruction (edge classification). One can import the
modules to convert the HEP data to different graph types, run the training, monitor the performance, and launch automatic hyperparameter tuning (missing for
now). It can be found at https://github.com/xju2/root_gnn
(documentations are in development). Below is a selection of its features/wishes:
- Common interface for converting physics events saved in different
data formats to (fully-connected) graph structures. To be developed
so as to allow different graph types (such as hypergraphs, customized
edges) - Pre-defined GNN models; Ultimately, we would like to cover all GNN models used in physics publications, i.e. one-stop GNN model shopping
station - Pytorch-lightning style trainers to make sure an easy training procedure
- Metrics monitoring
- Practical examples to get started with GNNs with public datasets
Author
Xiangyang Ju
(Lawrence Berkeley National Lab. (US))
