Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

Choose a workflow

The shortest route depends on the artifact you already have.

GoalWhat you needRecommended route
Check the installationA clean checkoutRun one quickstart smoke test
Understand the methodNothingRead Particle flow and MLPF
Evaluate a published modelA compatible checkpoint and ROOT fileDownload the model and use the detector-specific standalone evaluator
Train a modelA prepared TFDS datasetDownload one dataset split, then run mlpf ... train
Reproduce a datasetGenerator configuration and detector softwareRun simulation, postprocessing, validation, then TFDS conversion
Validate CMS reconstructionCMSSW outputUse the CMS validation workflow
Validate CLD or CLIC reconstructionEDM4hep ROOT plus a checkpointUse the Key4HEP standalone evaluator
Compare deployment formatsA checkpoint, model_kwargs.pkl, and TFDSRun PyTorch-to-ONNX numerical and timing validation

Three levels of use

1. Use published artifacts

This is the recommended starting point. Download a dataset or checkpoint and avoid detector simulation. It is the quickest route to model evaluation and training experiments.

2. Run a local smoke workflow

The scripts in scripts/local_test_*.sh download two small input files, build a small dataset, and run a short CPU training. They check that the software components fit together. They do not establish physics performance.

3. Produce data at scale

The Pixi and Snakemake workflow covers simulation, postprocessing, TFDS creation, training, and validation. A full campaign can use many batch jobs and substantial storage. It requires a reviewed site configuration and detector software access.

Detector choice

CLD and CLIC support both reconstructed track/cluster inputs and lower-level tracker/calorimeter hit inputs. The hit workflows have much larger events and are best treated as research workflows. See Current capabilities.