v1.7.2

• Added scikit-learn 1.8 compatibility.
• Removed debug print in RealMLP.
• fixed device memory estimation error in the scheduler when CUDA_VISIBLE_DEVICES was used.

v1.7.0

• added xRFM (D, HPO)
• added new 'tabarena-new' search space for RealMLP-HPO, including per-fold ensembling (more expensive)
  and tuning two more categorical hyperparameters
  (with better results)
• reduced RealMLP pickle size by not storing the dataset (#33)
• fixed gradient clipping for TabM
  (it did nothing previously, see #34).
  To ensure backward compatibility, it is set to None in the HPO search spaces now
  (it was already None in the default parameters).
• removed debug print in TabM training loop

v1.6.1

• For n_ens>1, changed the default behavior for classification to averaging probabilities instead of logits. This can be reverted by setting ens_av_before_softmax=True.
• Implemented time limit for HPO/ensemble methods through time_limit_s parameter.
• Support torch>=2.6 and Python 3.13.

v1.5.2

• v1.5.2: fixed more device bugs for HPO and ensembling

v1.5.1

• v1.5.1: fixed a device bug in TabM for GPU
• v1.5.0:
  • added n_repeats parameter to scikit-learn interfaces for repeated cross-validation
  • HPO sklearn interfaces (the ones using random search)
    can now do weighted ensembling instead by setting use_caruana_ensembling=True.
    Removed the RealMLP_Ensemble_Classifier and RealMLP_Ensemble_Regressor from v1.4.2
    since they are now redundant through this feature.
  • renamed space parameter of GBDT HPO interface
    to hpo_space_name so now it also works with non-TPE versions.
  • Added new TabArena search spaces for boosted trees (not TPE),
    which should be almost equivalent to the ones from TabArena
    except for the early stopping logic.
  • TabM now supports val_metric_name for early stopping on different metrics.
  • fixed issues #20 and #21 regarding HPO
  • small updates for the "Rethinking Early Stopping" paper

v1.4.2

• fixed handling of custom val_metric_name HPO models and Ensemble_TD_Regressor.
• If tmp_folder is specified in HPO models, save each model to disk immediately instead of holding all of them in memory. This can considerably reduce RAM/VRAM usage. In this case, pickled HPO models will still rely on the models stored in the tmp_folder.
• We now provide RealMLP_Ensemble_Classifier and RealMLP_Ensemble_Regressor, which will use weighted ensembling and usually perform better than HPO (but have slower inference time). We recommend using the new hpo_space_name='tabarena' for best results.

v1.4.0

What's Changed

• moved some imports to the new models optional dependencies to have a more light-weight RealMLP installation
• Added GPU support for CatBoost with help from Maximilian Schambach in #16 (not guaranteed to produce exactly the same results).
• Ensembling now saves models after training if a path is supplied, to reduce memory usage
• Added more search spaces
• fixed error in multiquantile output when the passed y was one-dimensional instead of having shape (n_samples, 1)
• Added some examples to the documentation

v1.3.0

• Added multiquantile regression for RealMLP: see the documentation
• More hyperparameters for RealMLP
• Added TabICL wrapper
• Small fixes