v2.0.0

DeepTab v2.0.0

DeepTab v2 is a major release. The whole package has been reorganized around a clean, sklearn-compatible split-config API, with a new internal layout, typed configuration objects, richer training and observability hooks, and a lightweight inference and serialization story.

If you are new to DeepTab, this is a great place to start. If you are coming from v1, please read the breaking change note below before you upgrade.

⚠️ Breaking change
The internal package layout, configuration objects, and import paths have all changed. Code written for v1 will not run unchanged. The Migration Guide walks you through every rename step by step.

Highlights

• Split-config API. Models are now configured through dedicated <Model>Config, PreprocessingConfig, and TrainerConfig objects instead of one long list of keyword arguments. Each config has a single, clear responsibility, so it is easier to read, reuse, and validate.
• New package structure. The library is split into clear modules: core, nn, architectures, configs, data, training, metrics, distributions, and hpo, each with its own public API.
• Stable and experimental split. Stable models are exported directly. Experimental models (ModernNCA, Tangos, Trompt) sit behind a lazy boundary so you always know what is production ready and what is still maturing.
• Lightweight inference wrapper. Save and load trained models through the .deeptab format, complete with rich serialization metadata for reproducible deployment.
• Observability. A new ObservabilityConfig is threaded through every estimator, with built-in support for experiment trackers.

New Features

• Optimizer and scheduler registry covering every torch.optim class, wired into both the Lightning module and TrainerConfig.
• A new deeptab.metrics registry for classification, regression, and distributional tasks that powers evaluate().
• More distributions, including Tweedie, inflated Poisson, and log-normal.
• Class-imbalance loss registry with weighted sampling.
• Model inspection API with a profile() dry run for pre-training diagnostics and read-only task_model access.
• Reproducibility helpers set_seed and seed_context.
• A typed exception hierarchy with sklearn-compatible error types.
• print_hardware_info for quick CPU, CUDA, and MPS detection.

Note for researchers
The metrics, distributions, and reproducibility helpers are designed to make experiments easier to seed, audit, and compare across runs.

Notable Fixes

• Stronger sklearn compatibility through check_is_fitted, __sklearn_is_fitted__, dataframe coercion, and ValueError-based error types.
• Transformer attention now uses batch_first to prevent cross-sample leakage.
• Custom torchmetrics are registered through nn.ModuleDict so their state moves with the model to the correct device.
• Distribution parameter transforms are applied before metric computation.
• DataLoader and sampler generators are now seeded for reproducible runs.

Migration

Every renamed config, import path, and the new save, load, and serve flow are documented in the Migration Guide. For the complete list of changes, see the CHANGELOG.md.

Pull Requests

• #400 DeepTab v2 API with split-config design (@mkumar73 )
• #401 Release v2.0.0 (@mkumar73 )

Full Changelog: v1.8.0...v2.0.0

v2.0.0rc2

Full Changelog: v2.0.0rc1...v2.0.0rc2

Full Changelog: v2.0.0rc1...v2.0.0rc2

v2.0.0rc1

What's Changed

• feat!: v2 release include split-config API, experimental models, and docs overhaul by @mkumar73 in #400

Full Changelog: v1.8.0...v2.0.0rc1

What's Changed

• feat!: v2 release include split-config API, experimental models, and docs overhaul by @mkumar73 in #400

Full Changelog: v1.8.0...v2.0.0rc1

v1.8.0

What's Changed

• feat: stable/experimental model split, save/load, and bug fixes by @mkumar73 in #346
• docs: sphinx upgrade 9.1.0 by @mkumar73 in #349
• Release/v1.8.0 by @mkumar73 in #351

Full Changelog: v1.7.0...v1.8.0

Release v1.7.0

What's Changed

Highlights

• Improved the testing framework with better cross-platform support.
• Fixed MambularDataset length handling for datasets with only categorical features.
• Simplified CI/CD and release workflows.
• Updated release process with improved PyPI publishing flow.
• Added documentation improvements and additional tests.
• Completed the final release bump from 1.7.0rc2 to 1.7.0.

Bug Fixes

• Fixed MambularDataset length for data with only categorical features by @MaxSchambach in #278.
• Merged 1.7.0 hotfixes back into main by @mkumar73 in #334.

CI/CD and Release Improvements

• Added conventional commit and semantic release setup by @mkumar73 in #295.
• Simplified branching, hardened release workflow, and updated PyPI publishing flow by @mkumar73 in #320.
• Fixed GitHub Actions workflow by simplifying package installation by @ChrisW09 in #286.
• Synced development branch after semantic release by @mkumar73 in #296.

Documentation, Testing, and Maintenance

• Enhanced testing framework with cross-platform support by @mhabedank in #266.
• Added release updates, documentation improvements, and tests by @mkumar73 in #326.
• Updated project configuration by @ChrisW09 in #274.

Repository Rename and Cleanup

• Updated repository/package references as part of the rename cleanup by @ChrisW09 in #282, #289, and #292.
• Merged development cleanup changes by @ChrisW09 in #284, #287, #290, and #293.

Release Preparation

• Prepared release candidate 1.7.0rc1 by @mkumar73 in #328.
• Finalized release 1.7.0 from 1.7.0rc2 by @mkumar73 in #335.

New Contributors

• @mhabedank made their first contribution in #266.
• @MaxSchambach made their first contribution in #278.

Full Changelog

v1.5.0...v1.7.0

What's Changed

• Enhanced Testing Framework with Cross-Platform Support by @mhabedank in #266
• Bump urllib3 from 2.3.0 to 2.5.0 by @dependabot[bot] in #273
• Bump requests from 2.32.3 to 2.32.4 by @dependabot[bot] in #272
• Bump torch from 2.5.1 to 2.7.0 by @dependabot[bot] in #270
• Update pyproject.toml by @ChrisW09 in #274
• Bump aiohttp from 3.11.13 to 3.12.14 by @dependabot[bot] in #276
• Fix MambularDataset length for data with only categorical features by @MaxSchambach in #278
• Feature/rename by @ChrisW09 in #282
• Develop by @ChrisW09 in #284
• Fix GitHub Actions workflow: simplify package installation by @ChrisW09 in #286
• Develop by @ChrisW09 in #287
• Feature/rename by @ChrisW09 in #289
• Develop by @ChrisW09 in #290
• Feature/rename by @ChrisW09 in #292
• Develop by @ChrisW09 in #293
• CI: Conventional commit and semantic release by @mkumar73 in #295
• ci: sync develop post semantic release by @mkumar73 in #296
• feat: conventional commit and semantic release by @mkumar73 in #306
• ci: simplify branching, harden release, update PyPI publishing flow by @mkumar73 in #320
• feat: release update, docs improvements, test added by @mkumar73 in #326
• bump: version 1.6.1 → 1.7.0rc1 by @mkumar73 in #328
• fix: merge 1.7.0 hotfixes into main by @mkumar73 in #334
• bump: version 1.7.0rc2 → 1.7.0 by @mkumar73 in #335

New Contributors

• @mhabedank made their first contribution in #266
• @MaxSchambach made their first contribution in #278

Full Changelog: v1.5.0...v1.7.0

v1.7.0rc2

What's Changed

• Enhanced Testing Framework with Cross-Platform Support by @mhabedank in #266
• Bump urllib3 from 2.3.0 to 2.5.0 by @dependabot[bot] in #273
• Bump requests from 2.32.3 to 2.32.4 by @dependabot[bot] in #272
• Bump torch from 2.5.1 to 2.7.0 by @dependabot[bot] in #270
• Update pyproject.toml by @ChrisW09 in #274
• Bump aiohttp from 3.11.13 to 3.12.14 by @dependabot[bot] in #276
• Fix MambularDataset length for data with only categorical features by @MaxSchambach in #278
• Feature/rename by @ChrisW09 in #282
• Develop by @ChrisW09 in #284
• Fix GitHub Actions workflow: simplify package installation by @ChrisW09 in #286
• Develop by @ChrisW09 in #287
• Feature/rename by @ChrisW09 in #289
• Develop by @ChrisW09 in #290
• Feature/rename by @ChrisW09 in #292
• Develop by @ChrisW09 in #293
• CI: Conventional commit and semantic release by @mkumar73 in #295
• ci: sync develop post semantic release by @mkumar73 in #296
• feat: conventional commit and semantic release by @mkumar73 in #306
• ci: simplify branching, harden release, update PyPI publishing flow by @mkumar73 in #320

New Contributors

• @mhabedank made their first contribution in #266
• @MaxSchambach made their first contribution in #278

Full Changelog: v1.5.0...v1.7.0rc2

Release v1.5.0

This release includes significant updates to the mambular package, focusing on integrating the pretab library for preprocessing, updating the version, and removing deprecated preprocessing modules. The most important changes include updating the documentation, modifying imports to use pretab, and removing old preprocessing code.

Documentation Updates:

• README.md: Updated preprocessing section to mention the use of pretab and provided links for further information.
• docs/api/preprocessing/Preprocessor.rst: Removed the Preprocessor class documentation.
• docs/api/preprocessing/index.rst: Removed the preprocessing module documentation.

Codebase Updates:

• mambular/__version__.py: Updated the version from 1.4.0 to 1.5.0.
• mambular/models/utils/sklearn_base_lss.py and mambular/models/utils/sklearn_parent.py: Changed imports from mambular.preprocessing to pretab.preprocessor and updated related code. [1] [2]

Removal of Deprecated Code:

• mambular/preprocessing/basis_expansion.py, mambular/preprocessing/ple_encoding.py, and mambular/preprocessing/__init__.py: Removed old preprocessing classes and methods. [1] [2] [3]

What's Changed

• change readme by @AnFreTh in #255
• TabR integration + compatibility check for PLR embedding with pre-processing types. by @bishnukhadka in #258
• merge master back to develop by @AnFreTh in #259
• Lss fix by @AnFreTh in #260
• v1.5.0 by @mkumar73 in #261
• preprocessor cleanup by @mkumar73 in #262
• v1.5.0 fixed by @mkumar73 in #263

New Contributors

• @bishnukhadka made their first contribution in #258

Full Changelog: v1.4.0...v1.5.0

Release 1.4.0

What's Changed

• Hotfix/docs by @mkumar73 in #247
• Modernnca by @AnFreTh in #251
  * Added MonderNCA and complete logic of using labels during training and inference into helper classes.
• Data check by @AnFreTh in #252
  * Introduced sanity checks in preprocessor module.
• Develop by @AnFreTh in #253

Full Changelog: v1.3.2...v1.4.0

Release 3.1.2

Fix error for binary classification loss function in sklearninterface

Release v1.3.1

This release (v1.3.1) introduces the new Tangos model to the Mambular library, along with updates to the documentation, versioning, and configuration files. The most important changes include adding the Tangos model, updating the version number, and modifying the lightning_wrapper.py and pretraining.py files to use the new model.

New Model Addition:

• mambular/base_models/tangos.py: Added the Tangos model, which is an MLP model with optional GLU activation, batch normalization, layer normalization, and dropout, including a penalty term for specialization and orthogonality.
• mambular/base_models/__init__.py: Included the Tangos model in the module imports and __all__ list.
• mambular/configs/tangos_config.py: Added configuration class for the Tangos model with predefined hyperparameters.

Documentation Updates:

• README.md: Updated to include the new Tangos model in the list of new models and added a description of the model. [1] [2]

Version Update:

• mambular/__version__.py: Incremented the version number from 1.3.0 to 1.3.1 to reflect the new changes.

Code Refactoring:

• mambular/base_models/utils/lightning_wrapper.py: Replaced instances of base_model with estimator to accommodate the new Tangos model. [1] [2] [3] [4] [5] [6] [7] [8] [9]
• mambular/base_models/utils/pretraining.py: Replaced instances of base_model with estimator to reflect the new model structure. [1] [2] [3] [4]

Configuration Updates:

• mambular/configs/__init__.py: Added DefaultTangosConfig to the module imports and __all__ list.
• mambular/models/__init__.py: Included TangosClassifier, TangosLSS, and TangosRegressor in the module imports and __all__ list.