Overview

Overview#

TabStruct is a modular framework for tabular machine learning with two pipelines:

  • Prediction: supervised learning on tabular data

  • Generation: synthesise tabular data and evaluate it

Key components#

  • Experiment runner (experiment/run_experiment.py): CLI entrypoint

  • Pipelines (experiment/pipeline): orchestrates model training/eval

  • Data layer (DataHelper, DataModule): split, curate, preprocess, and load data

  • Model layer (prediction/models, generation/models): sklearn baselines, Lightning models and other tabular models

  • Tuning (experiment/tune): Optuna sweeps

Metrics#

  • Classification: balanced_accuracy, F1_weighted, precision, recall, AUROC_weighted, ECE, cross_entropy_loss

  • Regression: rmse, mse, r2

W&B#

  • Project and entity are defined in src/tabstruct/common/__init__.py

  • All runs log summaries and artifacts under logs/ and W&B

Contents