quick start

mkdir cmsgnn 
cd cmsgnn 
git clone https://github.com/CMS-GNN-Tracking-Hackathon-2021/interaction-network

#create virtual env 
python3 -m venv cmsgnn_venv
source cmsgnn_venv/bin/activate

#install requirements 
pip install -r requirements.txt
pip install -r requirements-cpu-linux.txt
#pip install -r requirements-gpu-linux.txt

# build graph 
cd graph_construction 
# in some of the config files, you will have to change the directory to the eos space where the data is 
python build_graph.py config/select config file 

# run interaction network
python run_interaction_network.py --pt=2

interaction-network

This repo contains a version of the interaction network developed by Gage DeZoort, Savannah Thais et.al in their paper Charged particle tracking via edge-classifying interactionnetworks. This was created for the TrackML dataset, and thei repository is available here. The code in this repo is slightly simplified and adapted to CMSSW data. The main difference is an adaptation to fit the CMSSW Phase 2 geometry.

This repo is split into the following sections:

graph_construction

Builing the graphs. Beware that the parameters in the config files have been optimised for the TrackML data, not CMS data.

models

Code for the graph neural net

plotting

Evaluation plots

Eos space

The EOS space for the hackathon, with prepared ntuples, csv files, built graphs, trained neural nets: /eos/cms/store/group/ml/GNNTrackingHackathon

Building graphs in CMSSW

You can refer to this code by Levi Blinder as a starting point for porting the graph builidng to CMSSW. It does graph building in C++. https://github.com/leviBlinder/Graph_Construction_for_TrackML-C/blob/main/graph_construction/build_geometric.cc