Chen-Yifan / IN

Interaction Graph Network for Hbb

Interaction Network: a graph-based neural network for particle physics applications

The data used in this package was produced from CMS open simulated data using HiggsToBBNtupleProducerTool. The interaction network natively runs in PyTorch, but can be exported to other formats, as explained below.

E. Moreno et al., Interaction networks for the identification of boosted Higgs to bb decays, arXiv:1909.12285 [hep-ex]

Setup

Clone the repository and setup the libraries. Convert the data to h5 files which can be loaded through the H5Data class in data.py (TODO).

git clone https://github.com/eric-moreno/IN.git
cd IN
source install_miniconda.sh
source install.sh
source setup.sh # every time

Training

Change the test_path and train_path in IN_dataGenerator.py to reflect the directories of the test and training datasets (in converted h5 format). Change the file loading to reflect the naming scheme of your h5 files in IN_dataGenerator.py.

Determine the parameters needed for the IN. For example:

• Output directory = IN_training
• Vertex-particle branch = 1 (turned on)
• Vertex-vertex branch = 0 (turned off)
• De = 20
• Do = 24
• Hidden = 60

Would be executed by running:

python IN_dataGenerator.py IN_training 1 0 --De 20 --Do 24 --hidden 60 

For adversarial training to decorrelate from soft-drop jet mass, you must have a full pretrained IN to preload with the same parameters as the IN you are trying to adversarially train. For adversarial training you must specify the previous parameters as well as some additional parameters (preloaded IN directory, lambda weight, and mass bins):

python IN_dataGenerator_adv.py IN_training_adv 0 --De 20 --Do 24 --hidden 60 --preload IN_training --lambda 10 --nbins 40  

Alternatively, there is also an option to decorrelate using the DDT-technique. This is performed after training with a normal IN in the make_good_plots.py script.

Evaluation

Change the save path for the test dataset under save_path in IN_eval.py. Next call IN_eval.py given the network parameters and save location:

python IN_eval.py IN_training 1 0 --De 20 --Do 24 --hidden 60 

To make various plots (ROC, pT dep, PU dep, sculpting, distributions, etc.), run make_good_plots.py giving both the regular IN and mass decorrelated IN and an output directory:

xrdcp root://eosuser.cern.ch//eos/user/w/woodson/IN/output.pkl .
xrdcp root://eosuser.cern.ch//eos/user/w/woodson/IN/output_dec.pkl .
xrdcp root://eosuser.cern.ch//eos/user/w/woodson/IN/IN_training.tar.gz .
xrdcp root://eosuser.cern.ch//eos/user/w/woodson/IN/IN_training_adv.tar.gz .
tar xvzf IN_training_adv.tar.gz 
tar xvzf IN_training.tar.gz 
rm IN_training_adv.tar.gz IN_training.tar.gz
python make_good_plots.py IN_training IN_training_adv --outdir eval_IN_training 

This script (make_good_plots.py) will also create a DDT-version of the IN that decorrelates based on a mass-dependent threshold cut. It is usually better decorrelated than an adversarial training.

Exporting to TensorFlow, MXNet, etc.

This export uses ONNX, an open ecosystem for interchangable ML models. It works with both mass-sculpting and mass-decorrelated models. Change the save path for the test dataset under save_path in IN_onnx.py.

To use ONNX, you must have an already trained model and must provide IN_onnx.py with the parameters of this trained model:

python IN_onnx.py IN_training 1 0 --De 20 --Do 24 --hidden 60 

This will save the ONNX model in the directory where the trained IN is located, which can then be easily converted into most ML frameworks.