OpenTabular/DeepTab

What it solves

DeepTab simplifies the application of deep learning to tabular data. It removes the common friction points of manual preprocessing, architecture selection, and training loop boilerplate, providing a unified, scikit-learn compatible interface for a wide variety of neural network architectures designed specifically for tables.

How it works

DeepTab is built on PyTorch and PyTorch Lightning, wrapping complex neural architectures into a familiar fit/predict/evaluate workflow. It integrates with PreTab for automatic feature-type detection, encoding, and scaling. The library provides a "model zoo" of 15 stable architectures—including Mamba-inspired state space models, Transformers, and tree-inspired networks—each available as a classifier, regressor, or distributional (LSS) model for uncertainty quantification.

Who it’s for

It is designed for data scientists and ML engineers who want to leverage deep learning for tabular data without writing custom training loops or manual preprocessing pipelines, as well as researchers who need a standardized framework to benchmark different tabular deep learning architectures.

Highlights

• Scikit-learn API: Drops into existing pipelines and works with GridSearchCV for hyperparameter tuning.
• Broad Model Zoo: Includes 15 stable architectures across families like State Space Models (Mambular), Transformers (FTTransformer, TabTransformer), and Tree-inspired models (NODE).
• Automatic Preprocessing: Handles mixed feature types, scaling, and missing values automatically via PreTab.
• Distributional Regression: Supports LSS models that predict full probability distributions rather than just point estimates.
• Deployment-Ready: Features an InferenceModel for read-only prediction surfaces and a .deeptab format that bundles weights, schema, and preprocessing for easy reloading.
• Observability: Opt-in structured logging and integration with MLflow and TensorBoard for experiment tracking.