Harpoon

Harpoon is a framework for conditional tabular data generation and imputation using diffusion models with manifold-guided updates.

---

Repository Structure

The repository is organized to clearly separate training, sampling, and experiments:

• Training Scripts (train_*.py)
  Scripts for training models on different baselines and encodings (e.g., Harpoon, TabDDPM, GReaT, RePaint).

• Sampling Scripts (sampling_*.py)
  Scripts for generating samples or imputations using trained models. Includes general constraints, OHE/ordinal handling, and manifold-based updates.

• Experiments (experiments/)
  Stores experimental setups, batch scripts, and results from different runs.

• Datasets (datasets/)
  Scripts and folders for preparing and managing datasets, including baseline-specific subfolders.

• Visualization (visualization/)
  Manifold illustrations.

• Utility Scripts

  • utils.py – general-purpose utilities
  • dataset.py – dataset handling and constraints
  • diffusion_utils.py – sampling and model utilities
  • generate_mask.py – missing data masks
  • download_and_process.py – dataset download and preprocessing

• LaTeX Generators (latex_generator_*.py)
  Scripts for generating tables of experiment results.

• Tubular Neighbourhood Estimation (tubular_neighbourhood_estimator_*.py)
  Scripts for empirically estimating orthogonality.