v2.1.0

DeepTab 2.1.0

DeepTab 2.1 focuses on PreTab 1.0 integration, estimator reliability, persistence, training correctness, dependency security, and a stronger release pipeline.

This release also moves the supported Python range to 3.11–3.13 and includes fixes across preprocessing, metrics, HPO, retrieval models, LSS distributions, and model configuration.

Highlights

• Integrated PreTab 1.0 with preprocessing presets, a representation registry, canonical configuration, and dense/block output support.
• Improved save/load and estimator lifecycle reliability, including preprocessing state, loss functions, retrieval data, and architecture state.
• Fixed several silent correctness issues affecting training, metrics, HPO, LSS, TabR, ModernNCA, and architecture configuration.
• Strengthened CI/CD with release-branch CI, an aggregate CI Success gate, dependency endpoint testing, installed-wheel validation, and release-tag ancestry checks.
• Updated Python and dependency support to improve compatibility and security.

Added

• PreTab presets and a representation registry for selecting preprocessing strategies.
• Canonical PreprocessingConfig fields with legacy alias resolution and actionable errors for unsupported settings or missing optional dependencies.
• Block and dense output options in build_preprocessor.
• Saved preprocessor specifications, fingerprints, and feature widths for inspecting and restoring fitted preprocessing.

Changed

Python and dependency support

• DeepTab now requires Python 3.11–3.13 (>=3.11,<3.14). Python 3.10 reached end of life on 2026-10-01 and is no longer supported.
• Users requiring Python 3.10 should remain on DeepTab 2.0.x with compatible dependencies.
• Require pandas >=2.2.2,<3 and SciPy >=1.16,<2. The SciPy minimum avoids a reproduced native loader failure on Apple Silicon.
• Documentation builds require Python 3.12 or 3.13.

Preprocessing

• Integrated PreTab 1.0.
• New fits default numerical output_dim to 7; explicitly configured widths remain unchanged.
• Resolve preprocessing tasks automatically from estimator type.
• Propagate ObservabilityConfig.verbosity to preprocessing.
• Historical DeepTab 2.0 PLE artifacts require their original dependency environment and are not automatically converted.

Optional retrieval

• Moved FAISS and delu to the explicit tabr extra.
• The default installation and all observability bundle no longer install native retrieval packages.
• Native TabR remains unsupported on macOS arm64 and now fails before conflicting native imports.

Logging

• Model save/load status messages now use configured logging instead of print().

Fixed

Estimator state and persistence

• Preserve an existing fitted estimator when refit input validation fails, while correctly invalidating state when rebuilding has already started (#424, #426, #449).
• Apply set_params() overrides to the actual model configuration and reject unsupported parameters correctly (#410, #449).
• Refresh supplied data, validation splits, and loaders during rebuild=False continuation while retaining compatible weights and fitted preprocessing.
• Continue from pretrained weights by default in fit() (#446).
• Persist and restore fitted loss functions and NDTF's generated tree architecture (#444, #445).
• Default load() to CPU rather than automatically selecting hardware (#465).
• Preserve TabR and ModernNCA candidate data across save/load and loaded continuation.
• Prevent retrieval query-label leakage in TabR and ModernNCA (#413).
• Correct ModernNCA regression and binary-classification reload behavior and convert binary votes to real logits (#268).

Training, metrics, and HPO

• Preserve string class labels in Accuracy and F1Score, including explicit pos_label support for binary string labels.
• Respect fit-level arguments, optimizer settings, and custom regression losses instead of silently overriding them (#415, #430, #442).
• Respect checkpoint paths and requested monitor/mode settings, resolve best_epoch, and timestamp default checkpoint directories (#411, #425).
• Correct seeding, callbacks, fit-state transitions, and inference device handling (#452).
• Encode classification labels consistently and route metrics to labels, probabilities, or raw outputs as required (#409, #443).
• Aggregate training and validation metrics over complete epochs and correctly reset metric state.
• Correct expected calibration error handling and include confidence 1.0 in calibration bins (#421).
• Correct HPO trial selection, pruning, and winner refitting (#420).

Data and preprocessing

• Correct encoding output, categorical-feature metadata, input-dimension typing, and PreTab preprocessing compatibility (#422, #451).
• Cast concatenated features to floating point for non-embedding architectures.
• Normalize None to np.nan in object columns.
• Reject duplicate columns and correctly impute all-NaN columns.
• Reject categorical_method="none" with a clear configuration error.
• Correct regression target-shape handling (#441).

Architectures and distributions

• Correct architecture configuration options and neural-block behavior (#423, #450).
• Preserve strict custom architecture constructor signatures when supplying binary task metadata.
• Correct MLP bias handling and NDTF backward computation on MPS without changing saved parameter layouts.
• Apply the Mamba block correctly in MambaTab.forward (#435).
• Apply pre-normalization and dropout correctly around RowColTransformer attention (#436).
• Align contrastive pretraining anchors with their corresponding pairs (#437).
• Preserve parameter axes in ensemble LSS outputs (#447).
• Correct LSS scoring, distribution handling, and persistence (#419).
• Fix multinomial softmax, JohnsonSU naming, and singleton-dimension handling (#453).
• Use PyTorch's probability convention in NegativeBinomial (#429).
• Correct the inverted log-term sign in GammaDeviance (#427).

Security

• Require Torch >=2.13,<2.15.
• Require setuptools >=83,<84.
• Require Lightning >=2.6.6,<3.
• Require Accelerate >=1.15,<2.
• Require TensorBoard >=2.21,<3 for TensorBoard and tracking extras.
• Updated vulnerable locked dependencies.
• Removed blanket security-audit ignores.
• Updated the separately installed documentation HTML cleaner to 0.4.5.

CI, documentation, and maintenance

• Run CI automatically on release/** branches.
• Added the aggregate CI Success release and branch-protection gate.
• Validate release-tag ancestry before packaging, testing, and publication.
• Validate independently resolved minimum and newest supported dependency profiles.
• Test installed-wheel training and save/load workflows in CI.
• Correct CI dependency installation, runner-cache scope, and NumPy typing checks.
• Enable Prettier through pre-commit hooks and expand regression coverage.
• Clarify installation profiles, preprocessing widths, historical artifact limitations, supported environments, and the maintainer release process (#454).
• Centralize preprocessor construction and neural-block imports and remove unused helpers and stub modules.

Contributors

DeepTab 2.1.0 was primarily developed and coordinated by @mkumar73, with additional architecture, training, metric, and distribution fixes contributed by @ChrisW09.

Special thanks contributors:

• @ChrisW09 for MambaTab forward-path fixes (#435), contrastive pretraining alignment (#437), RowColTransformer attention fixes (#436), GammaDeviance correction (#427), and NegativeBinomial probability handling (#429).
• @mkumar73 for the broader DeepTab 2.1 integration and stabilization, including PreTab 1.0 integration, estimator lifecycle and persistence improvements, preprocessing, training, metrics, HPO, CI/CD, dependency compatibility, security updates, documentation, and release preparation (#477, #480).

Upgrade notes

Upgrade with:

pip install -U deeptab