root_gnn

GNN classification/Regression for reconstructed HEP events.

Install

conda create --name gnnana python=3.8
conda activate gnnana
conda install -c conda-forge jupyterlab
conda install -c conda-forge root

git clone https://github.com/xju2/root_gnn.git

cd root_gnn
pip install -e .

Input data

Download the W' and QCD data from https://zenodo.org/record/3981290#.XzQs5zVlAUFthe folder data, and split the W' data into wboson_big.txt and wboson_small.txt.

Commands for the edge classifier

• Graph Construction

create_tfrecord data/wboson_big.txt tfRec/fully_connected --max-evts 200 --evts-per-record 200 --type WTaggerDataset

The graphs are saved into the folder tfRec. Then split the folder into train, val, and test via the command:

split_files_for_nn tfRec inputs

• Graph Training

train_classifier train_wtaggers_edges.yaml

If train_wtaggers_edges.yaml does not exist in current folder, the program will look for the default one in the package, root_gnn/configs/train_wtaggers_edges.yaml.

• Evaluation

evaluate_wtagger data/wboson_small.txt train_wtaggers_edges.yaml test --nevts 10

• Metrics calculation

calculate_wtagger_metrics test.npz test

Commands for event classifier

Traing two event classifiers with different inputs, one from the edge classifier and the other from the anti-$k_t$ algorithm.

event classifier using outputs from the edge classifier

• Graph Construction for W boson events

create_tfrecord "tfRec/fully_connected*.tfrec" tfRec_filtered/wboson \
    --type WTaggerFilteredDataset \
    --signal --model-config configs/train_wtaggers_edges.yaml \
    --max-evts 100 --evts-per-record 100

• Graph Construction for q* events First create tfrecords for the edge classifier.

create_tfrecord data/qstar.txt tfrec_qcd/qcd --type WTaggerDataset --max-evts 100 --evts-per-record 100

Apply the edge classifier on the q* events

create_tfrecord "tfrec_qcd/qcd*.tfrec" \
	tfRec_filtered/qcd \
	--type WTaggerFilteredDataset  \
    --model-config configs/train_wtaggers_edges.yaml \
	--max-evts 100 --evts-per-record 100

Split these files

split_files_for_nn tfRec_filtered FilteredJets

• Training Global Classifier

train_classifier train_w_qcd.yaml

event classifier using outputs from the anti-$k_t$ algorithm

• Graph construction for W boson events

create_tfrecord data/wboson.txt \
	tfRec_ljet/wboson \
	--type WTaggerLeadingJetDataset \
	--signal \
	--max-evts 95000 --evts-per-record 1000

• Graphc osntruction for q* events

create_tfrecord data/qstar.txt \
	tfRec_ljet/qcd \
	--type WTaggerLeadingJetDataset \
	--max-evts 95000 --evts-per-record 1000

• Split the events

split_files_for_nn tfRec_ljet LeadingJets

• Graph training

train_classifier train_w_qcd_ljet.yaml

Evaluate both GNNs

evaluate_w_qcd_classifier "FilteredJets/test/*.tfrec" train_w_qcd.yaml classifier_gnn

evaluate_w_qcd_classifier "LeadingJets/test/*.tfrec" train_w_qcd_ljet.yaml classifier_ljet