tabular-GIA

Focused repo for the tabular gradient inversion (GIA) implementation and evaluation.

Setup

pip install -e .[dev]

Run the tabular GIA demo

python tabular_gia/main.py

This will load a checkpoint if present, otherwise train a model first.

To explicitly train:

python tabular_gia/train.py

Configuration

• tabular_gia/config.yaml controls dataset path, splits, batch size, and training settings.
• Datasets live under tabular_gia/data/... (already included).

Outputs

• Per-run metrics and per-row comparisons are written to tabular_gia/results.txt.
• Checkpoints are saved under tabular_gia/data/<dataset>/checkpoints/.

Notes

• The evaluation matches reconstructed rows to originals using Hungarian assignment before scoring.
• Numerical features are z-score normalized; categoricals are one-hot encoded.

Restoring upstream LeakPro

This repo intentionally prunes unrelated LeakPro components to keep the tabular GIA case small and reproducible. If you want the full LeakPro functionality later, you have two options:

1. Replace the leakpro/ folder with a fresh copy from the upstream LeakPro repo.
2. Add upstream LeakPro as a submodule (or dependency) and adjust imports accordingly.