Changes in version 0.8.0.9000                      

Bugfixes

  - Ancestor matrix is now taken into account for hierarchical
    classification (#188).

                 Changes in version 0.8.0 (2026-01-31)                  

New features

  - messaging is now improved with {cli}
  - add optimal threshold and support size into new 1.5 alpha entmax15()
    and sparsemax15() mask_types. Add an optional mask_topk config
    parameter. (#180)
  - optimizernow default to the torch_ignite_adam when available. Result
    is 30% faster pretraining and fitting tasks (#178).
  - add nn_aum_loss() function for area under the $Min(FPR,FNR)$
    optimization for cases of unbalanced binary classification (#178).
  - add a vignette on imbalanced binary classification with
    nn_aum_loss() (#178).

Bugfixes

  - config parameter now merge correctly for torch loss or torch
    optimizer generator.
  - nn_unsupervised_loss() is now a proper loss function.

                 Changes in version 0.7.0 (2025-04-16)                  

Bugfixes

  - Remove long-run example raising a Note.
  - fix tabet_pretrain failing with value_error("Can't convert data of
    class: 'NULL'") in R 4.5
  - fix tabet_pretrain wrongly used instead of tabnet_fit in Missing
    data predictor vignette
  - improve message related to case_weights not being used as
    predictors.
  - improve function documentation consistency before translation.
  - fix "..." is not an exported object from 'namespace:dials'" error
    when using tune() on tabnet parameters. (#160 @cphaarmeyer)

                 Changes in version 0.6.0 (2024-06-15)                  

New features

  - parsnip models now allow transparently passing case weights through
    workflows::add_case_weights() parameters (#151)
  - parsnip models now support tabnet_model and from_epoch parameters
    (#143)

Bugfixes

  - Adapt tune::finalize_workflow() test to {parsnip} v1.2 breaking
    change. (#155)
  - autoplot() now position the "has_checkpoint" points correctly when a
    tabnet_fit() is continuing a previous training using tabnet_model =.
    (#150)
  - Explicitely warn that tabnet_model option will not be used in
    tabnet_pretrain() tasks. (#150)

                 Changes in version 0.5.0 (2023-12-05)                  

New features

  - {tabnet} now allows hierarchical multi-label classification through
    {data.tree} hierarchical Node dataset. (#126)
  - tabnet_pretrain() now allows different GLU blocks in GLU layers in
    encoder and in decoder through the config() parameters
    num_idependant_decoder and num_shared_decoder (#129)
  - Add reduce_on_plateau as option for lr_scheduler at tabnet_config()
    (@SvenVw, #120)
  - use zeallot internally with %<-% for code readability (#133)
  - add FR translation (#131)

                 Changes in version 0.4.0 (2023-05-11)                  

New features

  - Add explicit legend in autoplot.tabnet_fit() (#67)
  - Improve unsupervised vignette content. (#67)
  - tabnet_pretrain() now allows missing values in predictors. (#68)
  - tabnet_explain() now works for tabnet_pretrain models. (#68)
  - Allow missing-values values in predictor for unsupervised training.
    (#68)
  - Improve performance of random_obfuscator() torch_nn module. (#68)
  - Add support for early stopping (#69)
  - tabnet_fit() and predict() now allow missing values in predictors.
    (#76)
  - tabnet_config() now supports a num_workers= parameters to control
    parallel dataloading (#83)
  - Add a vignette on missing data (#83)
  - tabnet_config() now has a flag skip_importance to skip calculating
    feature importance (@egillax, #91)
  - Export and document tabnet_nn
  - Added min_grid.tabnet method for tune (@cphaarmeyer, #107)
  - Added tabnet_explain() method for parsnip models (@cphaarmeyer,
    #108)
  - tabnet_fit() and predict() now allow multi-outcome, all numeric or
    all factors but not mixed. (#118)

Bugfixes

  - tabnet_explain() is now correctly handling missing values in
    predictors. (#77)
  - dataloader can now use num_workers>0 (#83)
  - new default values for batch_size and virtual_batch_size improves
    performance on mid-range devices.
  - add default engine="torch" to tabnet parsnip model (#114)
  - fix autoplot() warnings turned into errors with {ggplot2} v3.4
    (#113)

                 Changes in version 0.3.0 (2021-10-11)                  

  - Added an update method for tabnet models to allow the correct usage
    of finalize_workflow (#60).

                 Changes in version 0.2.0 (2021-06-22)                  

New features

  - Allow model fine-tuning through passing a pre-trained model to
    tabnet_fit() (@cregouby, #26)
  - Explicit error in case of missing values (@cregouby, #24)
  - Better handling of larger datasets when running tabnet_explain().
  - Add tabnet_pretrain() for unsupervised pretraining (@cregouby, #29)
  - Add autoplot() of model loss among epochs (@cregouby, #36)
  - Added a config argument to fit() / pretrain() so one can pass a
    pre-made config list. (#42)
  - In tabnet_config(), new mask_type option with entmax additional to
    default sparsemax (@cmcmaster1, #48)
  - In tabnet_config(), loss now also takes function (@cregouby, #55)

Bugfixes

  - Fixed bug in GPU training. (#22)
  - Fixed memory leaks when using custom autograd function.
  - Batch predictions to avoid OOM error.

Internal improvements

  - Added GPU CI. (#22)

                 Changes in version 0.1.0 (2021-01-14)                  

  - Added a NEWS.md file to track changes to the package.