pytorch_tabular

Why we included this project

Teams that have long defaulted to gradient boosting for structured data will find a real alternative here. PyTorch Tabular wraps a collection of deep architectures for tabular data behind one consistent, high-level API built on PyTorch Lightning, covering TabNet, TabTransformer, FT-Transformer, Neural Oblivious Decision Ensembles, and several newer attention- and gating-based models. You can benchmark a feed-forward net against a transformer-style model or a GATE variant by swapping a config instead of hand-writing each training loop. Categorical embedding, cross-validation folds, hyperparameter search, and model interpretability come as built-in workflows, and training scales from a single CPU to multiple GPUs without changing your code. Data scientists who want to test whether deep learning can beat their XGBoost baseline, along with researchers comparing tabular architectures, get the most value from the unified interface.