SPANet for up to 4 tops

1. Check out the GitHub repository

cd work
git clone https://github.com/tcoulvert/SPAtop/tree/topnet_dev

2. Install the Python venv (Ensure you have installed python 3.9 or greater)

python -m venv
pip install -e ./SPAtop

If for some reason the code does not run properly, it could be because packages may have changed and broken things. If that is the case, you can install your environment as follows:

python -m venv
pip install -r requirements.txt

3. Copy and convert the dataset(s)

Copy the Delphes ROOT TTree datasets from:

• LPC EOS: /eos/uscms/store/user/tsievert/ttbar_hadronic/ttbar_hadronic_*.root, or
• non-LPC EOS: root://cmseos.fnal.gov//store/user/tsievert/ttbar_hadronic/ttbar_hadronic_*.root

to the data/delphes/v1/ttbar_hadronic directory

Convert to training and testing HDF5 files.

python -m src.data.delphes.convert_to_h5 data/delphes/v1/ttbar_hadronic/sample_*.root --out-file data/delphes/v1/ttbar_hadronic_training.h5
python -m src.data.delphes.convert_to_h5 data/delphes/v1/ttbar_hadronic/sample_*.root --out-file data/delphes/v1/ttbar_hadronic_testing.h5

!!! WARNING !!! : From this step on, this repo hasn't been updated, so don't expect things to work. When the repo is updated, this README will change to reflect that.

5. Run the SPANet training

Override options file with --gpus 0 if no GPUs are available.

python -m spanet.train -of options_files/delphes/hhh_v2.json [--gpus 0]

SPAtop Training via Kubernetes

Prerequisites

Training via kubernetes on the cms-ml namespace requires the following:

• kubectl configured to target the cms-ml namespace
• PersistentVolumeClaim named spatopvol containing training data (already created)
• Docker image gitlab-registry.nrp-nautilus.io/jmduarte/hhh:latest (to be updated)

---

Data Organization

Data should be placed under the PVC at:

spatopvol/data/delphes/v1/tt_training.h5

---

Configuration Files

SPAtop training requires the following files, each located within the PVC:

File	Description	Location
tt_hadronic.yml	Physics process settings	/spatopvol/event_files/tt_hadronic.yml
spatop_v1.json	SPANet model parameters	/spatopvol/options_files/spatop_v1.json

Additionally, the Kubernetes job manifest is required:

• spatop-job-train.yml: Defines the Job spec for launching the SPAtop training container. Place this file in your local working directory where you run kubectl commands.

---

Launching Training

To start the SPAtop training job, apply the Kubernetes manifest:

kubectl apply -f spatop-job-train.yml -n cms-ml

This command creates a Kubernetes Job that spawns one or more pods to perform the training.

---

Monitoring Jobs

1. List Jobs and Pods

   kubectl get jobs -n cms-ml
   kubectl get pods -l job-name=spatop-job-train -n cms-ml

2. Describe a Pod

   kubectl describe pod <pod-name> -n cms-ml

3. Stream Logs

   kubectl logs -f <pod-name> -n cms-ml

Repeat these commands to track pod status, resource usage, and training progress.

---

Cleanup

Once training completes successfully,remove the job and its pods:

kubectl delete job spatop-job-train -n cms-ml

---

6. Evaluate the SPANet training

Assuming the output log directory is spanet_output/version_0. Add --gpu if a GPU is available.

python -m spanet.test spanet_output/version_0 -tf data/delphes/v2/hhh_testing.h5 [--gpu]

7. Evaluate the baseline method

python -m src.models.test_baseline --test-file data/delphes/v2/hhh_testing.h5

Instructions for CMS data set baseline

The CMS dataset was updated to run with the v26 setup (nAK4 >= 4 and HLT selection). The update includes the possibility to apply the b-jet energy correction. By keeping events with at a least 4 jets, the boosted training can be performed on a maximum number of events and topologies.

List of samples (currently setup validated using 2018):

/eos/user/m/mstamenk/CxAOD31run/hhh-6b/cms-samples-spanet/v26/GluGluToHHHTo6B_SM_spanet_v26_2016APV.root
/eos/user/m/mstamenk/CxAOD31run/hhh-6b/cms-samples-spanet/v26/GluGluToHHHTo6B_SM_spanet_v26_2016.root
/eos/user/m/mstamenk/CxAOD31run/hhh-6b/cms-samples-spanet/v26/GluGluToHHHTo6B_SM_spanet_v26_2017.root
/eos/user/m/mstamenk/CxAOD31run/hhh-6b/cms-samples-spanet/v26/GluGluToHHHTo6B_SM_spanet_v26_2018.root

To run the framework, first convert the samples (this will allow to use both jets pt or ptcorr, steerable from the configuration file:

mkdir data/cms/v26/
python -m src.data.cms.convert_to_h5 /eos/user/m/mstamenk/CxAOD31run/hhh-6b/cms-samples-spanet/v26/GluGluToHHHTo6B_SM_spanet_v26_2018.root --out-file data/cms/v26/hhh_training.h5
python -m src.data.cms.convert_to_h5 /eos/user/m/mstamenk/CxAOD31run/hhh-6b/cms-samples-spanet/v26/GluGluToHHHTo6B_SM_spanet_v26_2018.root --out-file data/cms/v26/hhh_testing.h5

Then training can be done via:

python -m spanet.train -of options_files/cms/hhh_v26.json --gpus 1

Two config files exist for the event options:

event_files/cms/hhh.yaml # regular jet pT
event_files/cms/hhh_bregcorr.yaml # jet pT with b-jet energy correction scale factors applied

Note: to run the training with the b-jet energy correction applied, the log_normalize of the input variable was removed. Keeping it caused a 'Assignement collision'.