SimplJets

A SNN built on top of TopModelX library for jet classification on jetnet dataset

Example Usage

To do the training run yourself, use a high RAM CPU or GPU's preferrably (i used 8x H100's for this)

for non GPU run -> simplical.py for a GPU run -> simplicial_gpu.py

cd simplicialjets
python3 -m venv venv 
source venv/bin/activate
pip3 install torch_geometric networkx JetNet tqdm toponetx topomodelx
python3 simplicial.py
#or
python3 simplicial_gpu.py

Working

1. Imports the jetnet dataset with the types of jets as gluons, top-quarks and W bosons and features as

   a. eta_rel = relative psuedorapidity -> gives a measure of how spread/clustered the jet particles are along the length or eta.

   b. psi_rel = relative azimuthal angle -> gives measure of how spread/clustered the jet particles are along the width or beam axis

   c. pt_rel = relative transverse momentum -> intuitively: how close a particle is from the axis or in any angle way

   d. mask = indicator of presence/absence of particle: shows if the particle is present:1 otherwise 0 e: shows if the particle is present:1 otherwise 0 e: shows if the particle is present:1 otherwise 0 e: shows if the particle is present:1 otherwise 0

2. Constructs the graph from the data points with the KNN graphs

3. These are then lifted into a higher dimension by face construction by finding cliques that have length=3 and setting seperate features

4. Find the up,down and hodge laplacians at the edge level and get the features for each graph

5. Define the arch :

  (base_model): SCNN(
    (layers): ModuleList(
      (0-1): 2 x SCNNLayer()
    )
  )
  (linear): Linear(in_features=16, out_features=3, bias=True)

6. Train it!

config

you can just tweak anything you want pretty easily for more stuff to add also checkout the topomodel docs

maybe change it completely from a SNN?

go forth and explore!