You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
Add presets to TabPFNClassifier and TabPFNRegressor, named bundles of inference_config values. TabPFNRegressor(presets="zero_inflated") is for targets with many exact zeros, and "grouped" pools the predictions of a group. Presets combine, as in ["zero_inflated", "grouped"]. The classifier offers only "grouped". (#453)
TabPFNClassifier and TabPFNRegressor gain cross_val_predict_splits(), which fits and predicts splits the caller built, each with its own rows, in one cross-validation job on the server, so each split's features can come from preprocessing fitted on that split's training rows only. A job it stopped waiting for is resumed with wait_for_cross_val_predict_splits(). (#456)
Changed
Accept average_before_softmax="auto", which takes the setting of the checkpoint, as the default does. (#453)
Allow group_col without thinking mode: the client sends it to the server in tabpfn_config, and the model replaces the group columns with one categorical feature. (#454)
Fixed
Training sets with fewer than 2 rows now raise a clear ValueError before uploading, in the hosted, Azure AI Foundry and SageMaker estimators as well, instead of failing on the server with a 500. (#447)