Recipe for FastPUPPI framework

cmsrel CMSSW_15_1_0_pre4
cd CMSSW_15_1_0_pre4/src
cmsenv
git cms-init
git cms-addpkg DataFormats/L1TParticleFlow
git cms-addpkg DataFormats/L1TCorrelator
git cms-addpkg L1Trigger/Phase2L1ParticleFlow
git cms-addpkg L1Trigger/TrackTrigger
git cms-addpkg SimTracker/TrackTriggerAssociation
git cms-addpkg L1Trigger/VertexFinder

git cms-checkout-topic -u friti:new_G_branches_and_nnvtx_on

git clone git@github.com:elfontan/FastPUPPI.git -b puppiML_target

It includes

• puppi weight and NNVtx score in the PUPPI collection
• NNVtx association ON for the charged tracks
• PUPPI collection is the same as PF candidates (useful for PUPPI studies)
• New sums and ratios are saved in the ntuples for a dR of 0.2

Notebooks

MET computation

• The notebook Reproduce_MET.ipynb includes code to check that the MET and PUPPI computations on top of fastPUPPI ntuples is correct and it works.
• The notebook MET_pipeline.ipynb runs on fastPUPPI ntuples and computes the new PUPPi collection with new targets, compares it with the original PUPPI collection, and computes the MET for both.

Jet validation

• The notebook reclusteringSCJets.ipynb includes code to perform a reclustering to produce seeded-cone jets starting from two different collections

CoffeaForFastPUPPI

Run coffea on the ntuples

Last update: 2026 April 7th

• The coffea framework will read the FastPUPPI ntuples and make cuts & plot stuff.
• Plan is to update cut & histogramming configs to streamline a bit better, but here's a working version.

Workflow

1. Open runAnalyzer_pf.py (or for background, runAnalyzer_bkg.py).

• Add the input file location or the input file itself. For example, you can put the location of the perfNano.root files or the hadded root file itself.
• Update the output coffea file name.
• Choose histo_config and cut_config

2. For histo_config: it's a bit messy but choose the yaml file and update it per needed.

3. For cut_config: most of functions needed for applying relevant cuts are implemented in cut_config_pf.py (KP: let me know if more features are needed). Change the bottom of the file CUTFLOW to modify any cuts to be applied. Top of the file lists some cut values.

4. Run python3 runAnalyzer_pf.py.

5. Plotting & cutflow related scripts can be found in plotting repo. (TO-DO: Add example of these codes)

Training

(Preliminary) GNN training for regression

Preliminary training on GNN (EdgeConv, GravNet) can be found in the Training repo.

1. Set up environment on lxplus:

   ssh <your_username>@lxplus-gpu.cern.ch
   
   apptainer shell -B /afs -B /eos -B /eos/user/k/kypark -B /eos/home-k/kypark -B /afs/cern.ch/user/k/kypark -B /etc/sysconfig/ngbauth-submit -B ${XDG_RUNTIME_DIR} --env KRB5CCNAME="FILE:${XDG_RUNTIME_DIR}/krb5cc" --nv /cvmfs/unpacked.cern.ch/registry.hub.docker.com/cmsml/cmsml:latest --bind /usr/local/cuda-12.4:/usr/local/cuda-12.4
   
   python -m venv --system-site-packages bestpuppi
   source bestpuppi/bin/activate
   export CUDA_HOME=/usr/local/cuda-12.4
   
   pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv torch_geometric -f https://data.pyg.org/whl/torch-2.6.0+cu124.html
   
   

2. Prepare dataset

   # Example
   python3 prepare_dataset_batches.py --input perfNano_TTbar_PU200.root --output_dir ./graphs_dr0p3 --graphs_per_file 50000
   
   # Example for dR < 0.2 dataset
   python3 prepare_dataset_batches.py --input perfNano_TTbar_PU200.root --dr 0.2 --output_dir ./graphs_dr0p2 --graphs_per_file 50000
   
   

   • To add more inputs, change this script (once the branches are available on the ntuples)
   • Likewise, you can turn off some inputs from the script
   • Some pre-processed inputs (50000 neutral cones per pt file) with input feature metadata can be found here: /eos/cms/store/group/cmst3/group/l1tr/kypark/puppi_training_neutral_dr0p3_includeTrk/

   2.1 Print some dataset

   # update the pt file path in the python script and run:
   python3 print_dataset.py

   This will print out the current dataset metadata and the input feature for one neutral

   2.2 Dataset validation (plot the input features from the dataset)

   # Example                         
   python3 validate_dataset.py \
     --input "./graphs/*.pt" \
     --output_dir ./plots \                 
     --max_files 10                         

3. Run training

   # DeepSet (baseline for now)
   python train_deepset.py \
     --data "./graphs/*.pt" \
     --epochs 30 \
     --output_activation softplus \
     --loss_type weighted_huber \
     --penalty_alpha 1 \
     --feature_importance
   
   # Example (for GNN)
   python3 train_gravnet.py --data "./graphs/*.pt" --epochs 20 --batch_size 128
   
   

   It will run the training and save at every epoch the training output & model at runs/run_*_* directory.

4. Check training: plot loss history, regression vs. target, residual plots, etc

   python3 plot_eval.py --run runs/run_*_*