CHANGELOG.md

1.0.3

• Fix another bug in DDUMixin, similar to the one in version 1.0.2.

1.0.2

• Fix bug in DDUMixin that called an outdated function that was renamed in 1.0.0.

1.0.0

• Massive changes to repository structure, including restructuing of APIs and bug fixes. See the points below:
  • Demo notebook added, showcasing some common use cases. Take a look here or here.
  • API changes:
    • Instead of using a model_params dict when initializing a Model subclass, the necessary model arguments are spelled out explicitly, similarly to the API of scikit-learn` models. Furthermore, reasonable default values are set for most model parameters.
    • The Module class implements additional functionalities:
      • An available_uncertainty_metrics attribute gives some information about available uncertainty metrics, mapping from their names to their functions.
      • get_sequence_representation() was renamed to get_sequence_representation_from_hidden(). This was done to distinguish it from the new get_sequence_classification() function, which retrieves sequence representation directly from the inputs.
      • get_hidden_representation() defines how to obtain hidden representations from an input.
    • An explicit function called compute_loss_weights() was added to the Model class to allow to customize loss weights for unbalanced problems.
    • BERT-related models now have a bert_class argument, that allow the underlying model to be changed to a RoBERTa, DistilBert or similar model.'
    • Split utils.uncertainty_eval into utils.uncertainty_eval and utils.calibration_eval.
    • There are now explicit functions to evaluate calibration and uncertainty properties on model, specified in utils.uncertainty_eval.evaluate_uncertainty() and utils.calibration_eval.evaluate_calibration().
  • Bug fixes:
    • Changing batch size during inference won't produce shape errors for LSTM models anymore.

0.9.2

• Fix import issues concerning the transformers and protobuf packages mentioned in issues #8 and #9.