v2.0.0.dev1

This intermediate release serves as a preliminary version of the upcoming v2.0.0. Consider it an alpha release, where interface changes are still possible.

Added

• General
  • Python requirements raised to Python 3.8 since Python 3.7 has reached end of life on 2023-06-27.
  • Dropped torchtext as an integration dependency. For individual use cases it can of course still be used.
  • Added environment variables SMALL_TEXT_PROGRESS_BARS and SMALL_TEXT_OFFLINE to control the default behavior for progress bars and model downloading.
• PoolBasedActiveLearner:
  • initialize_data() has been replaced by initialize() which can now also be used to provide an initial model in cold start scenarios. (#10)
• Classification:
  • All PyTorch-classifiers (KimCNN, TransformerBasedClassification, SetFitClassification) now support torch.compile() which can be enabled on demand. (Requires PyTorch >= 2.0.0).
  • All PyTorch-classifiers (KimCNN, TransformerBasedClassification, SetFitClassification) now support Automatic Mixed Precision.
  • SetFitClassification.__init__() now has a verbosity parameter (similar to TransformerBasedClassification) through which you can control the progress bar output of SetFitClassification.fit().
  • TransformerBasedClassification:
    • Removed unnecessary token_type_ids keyword argument in model call.
    • Additional keyword args for config, tokenizer, and model can now be configured.
• Embeddings:
  • Prevented unnecessary gradient computations for some embedding types and unified code structure.
• Pytorch:
  • Added an inference_mode() context manager that applies torch.inference_mode or torch.no_grad for older Pytorch versions.
• Query Strategies:
  • New strategies: DiscriminativeRepresentationLearning, LabelCardinalityInconsistency, ClassBalancer, and ProbCover.
  • Query strategies now have a tie-breaking mechanism to randomly permutate when there is a tie in scores.
  • Added ScoringMixin to enable a reusable scoring mechanism for query strategies.
  • LightweightCoreset can now process input in batches. (#23)
• Vector Index Functionality:
  • A new vector index API provides implementations over a unified interface to use different implementations for k-nearest neighbor search.
  • Existing strategies that used a hard-coded vector search ([ContrastiveActiveLearning][contrastive_active_learning], [SEALS][seals], [AnchorSubsampling][anchor_subsampling]) have been adapted and can now be used with different vector index implementations.

Fixed

• Fixed a bug where the clone() operation wrapped the labels, which then raised an error. This affected the single-label scenario for PytorchTextClassificationDataset and TransformersDataset. (#35)
• Fixed a bug where the batching in greedy_coreset() and lightweight_coreset() resulted in incorrect batch sizes. (#50)
• Fixed a bug where lightweight_coreset() failed when computing the norm of the elementwise mean vector.

Changed

• General
  • Moved split_data() method from small_text.data.datasets to small_text.data.splits.
• Dependencies
  • Raised setfit version to 1.1.0.
• Classification:
  • The initialize() methods of all PyTorch-classifiers (KimCNN, TransformerBasedClassification, SetFitClassification) are now more unified. (#57)
  • KimCNNClassifier / TransformerBasedClassification: model selection is now disabled by default. Also, it no longer saves models when disabled, thereby greatly reducing the runtime.
• Utils
  • init_kmeans_plusplus_safe() now supports weighted kmeans++ initialization for scikit-learn>=1.3.0.

Removed

• Deprecated functionality
  • Removed default_tensor_type() method.
  • Removed small_text.utils.labels.get_flattened_unique_labels().
  • Removed small_text.integrations.pytorch.utils.labels.get_flattened_unique_labels().
  • Classification
    • Removed early stopping legacy arguments in __init__() for KimCNN and TransformerBasedClassification. (Use fit() keyword arguments instead.)
    • Removed model selection legacy argument in TransformerBasedClassification.__init__().
• The explicit installation instruction for conda was removed, but the small-text conda-forge package will remain.

v1.4.1

Bugfix release.

Fixed

• Fixed an out of bounds error that occurred when DiscriminativeActiveLearning queries all remaining unlabeled data.
• Fixed typos/wording in PoolBasedActiveLearner docstrings.
• Pinned SetFit version in notebook example. (#64)
• Fixed an out of bounds error that could occur in SetFitClassification for both 32bit systems and Windows. (#66)
• Fixed errors in notebook examples that occurred with more recent seaborn / matplotlib versions.

Changed

• Documentation: added links to bibliography. (#65)

v1.4.0

Fixes SetFit seed control and adds the AnchorSubsampling query strategy.

Added

• New query strategy: AnchorSubsampling.

Fixed

• Changed the way how the seed is controlled in SetFitClassification since the seed was fixed unless explicitly set via the respective trainer keyword argument.

Changed

• Documentation: Added a section where compatible transformer models are listed.
• Documentation: Updated showcase section.

v1.3.3

Bugfix release.

Changed

• An errata section was added to the documentation.

Fixed

• Fixed a deviation from the paper, where DeltaFScore also considered negative label predictions for the agreement. (#51)
• Fixed a bug in KappaAverage that affected the stopping behavior. (#52)

Contributors

@zakih2 @vmanc

v1.3.2

Bugfix release.

Fixed

• Fixed a bug in TransformerBasedClassification where validations_per_epoch>=2 left the model in eval mode. (#40)

v1.3.1

Bugfix release.

Fixed

• Fixed a bug where parameter groups were omitted when using TransformerBasedClassification's layer-specific fine-tuning functionality. (#36, #38)
• Fixed a bug where class weighting resulted in nan values. (#39)

Contributors

@JP-SystemsX

v1.3.0

SetFitClassification now also supports dropout sampling (like KimCNNClassifier and TransformerBasedClassification).

Added

• Added dropout sampling to SetFitClassification.

Fixed

• Fixed broken link in README.md.
• Fixed typo in README.md. (#26)

Changed

Stopping Criteria

• The ClassificationChange stopping criterion now supports multi-label classification.

Documentation

• Updated the active learning setup figure.
• The documentation of integrations has been reorganized.

Contributors

@rmitsch

v1.2.0

This release adds a SetFit classifier, the BALD query strategy, and two new example notebooks.

Added

Active Learning

• PoolBasedActiveLearner now handles keyword arguments passed to the classifier's fit() during the update() step.
• New strategy: BALD.
• SubsamplingQueryStrategy now uses the remaining unlabeled pool when more samples are requested than are available.

Classification

• Added new classifier: SetFitClassification which wraps huggingface/setfit.

Examples

• Revised both existing notebook examples.
• Added a notebook example for active learning with SetFit classifiers.
• Added a notebook example for cold start initialization with SetFit classifiers.

Documentation

• A showcase section has been added to the documentation.

Fixed

• Distances in lightweight_coreset were not correctly projected onto the [0, 1] interval (but ranking was unaffected).

Changed

• Coreset implementations now use the distance-based (as opposed to the similarity-based) formulation.

v1.1.1

Minor bug fix release.

Fixed

• Fixed model selection which could raise an error under certain circumstances (#21).

v1.1.0

This release adds a conda package, more convenient imports, and improves many aspects of the classifcation functionality. Moreover, one new query strategy and three stopping criteria have been added.

Added

General

• Small-Text package is now available via conda-forge.
• Imports have been reorganized. You can import all public classes and methods from the top-level package (small_text):

  from small_text import PoolBasedActiveLearner
  

Classification

• All classifiers now support weighting of training samples.
• Early stopping has been reworked, improved, and documented (#18).
• Model selection has been reworked and documented.
• [!] KimCNNClassifier.__init()__: The default value of the (now deprecated) keyword argument early_stopping_acc has been changed from 0.98 to -1 in order to match TransformerBasedClassification.
• [!] Removed weight renormalization after gradient clipping.

Datasets

• The target_labels keyword argument in __init()__ will now raise a warning if not passed.
• Added from_arrays() to SklearnDataset, PytorchTextClassificationDataset, and TransformersDataset to construct datasets more conveniently.

Query Strategies

• New multi-label strategy: CategoryVectorInconsistencyAndRanking.

Stopping Criteria

• New stopping criteria: ClassificationChange, OverallUncertainty, and MaxIterations.

Deprecated

• small_text.integrations.pytorch.utils.misc.default_tensor_type() is deprecated without replacement (#2).
• TransformerBasedClassification and KimCNNClassifier:
  The keyword arguments for early stopping (early_stopping / early_stopping_no_improvement, early_stopping_acc) that are passed to __init__() are now deprecated. Use the early_stopping
  keyword argument in the fit() method instead (#18).

Fixed

Classification

• KimCNNClassifier.fit() and TransformerBasedClassification.fit() now correctly
  process the scheduler keyword argument (#16).

Removed

• Removed the strict check that every target label has to occur in the training data.
  (This is intended for multi-label settings with many labels; apart from that it is still recommended to make sure that all labels occur.)