This is the widely used PyTorch implementation of TabNet, Google’s attentive deep learning architecture for tabular data, maintained by DreamQuark AI. It brings TabNet’s signature combination โ deep learning accuracy with built-in feature-selection interpretability โ to scikit-learn users through familiar classifier and regressor classes. The repository has nearly 3,000 stars on GitHub and underpins the popular pytorch-tabnet package.
Tabular data is where gradient boosting traditionally beats neural networks, partly because trees handle heterogeneous features gracefully and partly because they are interpretable. TabNet narrows that gap: its sequential attention mechanism picks which features to reason with at each decision step, learning sparse feature masks that double as explanations of the model’s behavior. This implementation wraps the architecture in a clean, production-friendly API with support for classification, regression, multi-task learning and unsupervised self-supervised pretraining on unlabeled tables. It handles details like categorical embeddings, missing values and GPU training out of the box. The project provides extensive documentation, worked examples on well-known datasets, and visualizations of the feature masks, which makes it straightforward to present model explanations to non-technical audiences. It integrates cleanly with standard scikit-learn workflows such as cross-validation and hyperparameter search.
The library is completely free and open source under the MIT license. There is no commercial edition โ install it from PyPI and train on your own infrastructure.
This TabNet implementation is for data scientists working with structured tables โ credit scoring, churn, fraud, forecasting โ who want deep learning performance while retaining feature-level explanations for stakeholders and regulators.