PyTorch workflow

The PyTorch code under pytorch_src/ provides reusable datasets, graph layers, models, trainers, and inference managers. Configuration-driven entry points in scripts/ assemble these components into a complete experiment.

Recommended procedure

1. Set up Python. Follow the installation guide and verify that PyTorch can see the intended CPU or GPU device.
2. Prepare and inspect data. Read Datasets and data loaders to choose an on-disk format, create train/validation loaders, and understand the graph-batch fields.
3. Select or implement a model. Follow Writing a model for the input/output contract, configuration, checkpoint metadata, and a forward-pass smoke test.
4. Configure and run training. Follow Training to write a trainer, create the JSON configuration, run debug mode, save checkpoints, and monitor TensorBoard.
5. Evaluate a checkpoint. Follow Inference to implement an inference manager, tune clustering thresholds, and generate metrics and diagnostic plots.

In short:

converted data -> dataset/loader -> model -> trainer -> checkpoint -> inference/report

Main components

Component	Location	Purpose
Dataset and loader	pytorch_src/datasets/nps.py	Read event files, collate graphs, and split validation data
Models	pytorch_src/models/	Map graph features to object-condensation predictions
Trainers	pytorch_src/training/	Optimize models, validate, log, checkpoint, and export ONNX
Inference	pytorch_src/inference/	Restore checkpoints, cluster predictions, calculate metrics, and report
Configuration	pytorch_src/utils/config.py	Dynamically construct components from JSON
Training entry point	scripts/train.py	Run a configured training experiment
Inference entry point	scripts/inference.py	Evaluate a configured checkpoint

The tutorials use the hit-level object-condensation pipeline as the concrete example, but the same component boundaries support custom models, trainers, and inference logic.

Use the implementation checklist before starting a full training run or handing a model off for deployment.