v1.2.0

TabDPT v1.2

We're pleased to announce TabDPT v1.2, or TabDPT-Turbo! This release accelerates TabDPT fitting and inference by multiple orders of magnitude (~120x on average on TabArena) while improving predictive performance.

Changes:

• Long-context training and inference. context_size is now set to None by default, meaning no context limit, which is usually optimal if your memory can accommodate it. Set this value lower to control GPU memory.
• No retrieval by default. This can be reenabled using the context_reduction parameter, but tends to be much slower. If needed for memory reasons, we recommend trying context_reduction="subsample" or context_reduction="subsample-balanced" for imbalanced problems, and increasing n_ensembles to recover performance.
• inf_batch_size is now a predict function parameter called batch_size and has a higher default value. This parameter can also help control GPU memory usage.
• New model weights are now used, available on Huggingface and automatically downloaded by default.

Paper link coming soon.

Code Changes

• TabDPT v1.2 by @j201 in #74

Full Changelog: v1.1.14...v1.2.0

Contributors

@j201 @rasaHosseinzadeh @Barry0121 @ikozl090