Bayesflow

6. Publication-ready calibration diagnostic for expected calibration error (ECE) in a model comparison setting has been
added to ``diagnostics.py`` and is accessible as ``plot_calibration_curves()``
7. A new module ``experimental`` has been added currently containing ``rectifiers.py``.
8. Default settings for transformer-based architectures.
9. Numerical calibration error using ``posterior_calibration_error()``

General Improvements:
1. Improved docstrings and consistent use of keyword arguments vs. configuration dictionaries
2. Increased focus on transformer-based architectures as summary networks
3. Figures resulting ``diagnostics.py`` have been improved and prettified
4. Added a module ``sensitivity.py`` for testing the sensitivity of neural approximators to model misspecification
5. Multiple bugfixes, including a major bug affecting the saving and loading of learnable permutations

----------
1. Fix bug failing to propagate global context variables for model comparison.
2. Major revamp of tutorials.
3. Update dependencies and continuous integration.

----------

1. Add ``bidirectional`` flag to ``SequentialNetwork`` and ``TimeSeriesTransformer`` for potential to improve
performance.
2. Deprecate name ``SequentialNetwork`` and use ``SequenceNetwork`` instead to avoid confusion with ``tf.keras.Sequential``.
3. Change default to ``use_layer_norm=False`` of ``SetTransformer`` due to superior performance on relevant exchangeable models.

----------

1. Bugfix in ``SimulationMemory`` affecting the use of empty folders for initializing a ``Trainer``
2. Bugfix in ``Trainer.train_from_presimulation()`` for model comparison tasks
3. Added a classifier two-sample test function ``c2st`` in ``computational_utilities``

----------

1. Bugfix in ``SetTransformer`` affecting saving and loading when using the version with inducing points.
2. Bugfix in ``SetTransformer`` when using ``train_offline`` and batches result in unequal shapes.
3. Improved documentation with examples

----------

Major (Breaking)

1. Coupling layers have been refactored to ensure easy interoperability between spline flows and affine coupling flows
2. New internal classes and layers have been added! Saving and loading of old models will not work! However, the interface
remains consistent.
3. Model comparison now works for both hierarchical and non-hierarchical Bayesian models. Classes have been generalized
and semantics go beyond the ``EvidentialNetwork``
4. Default settings have been changed to reflect recent insights into better hyperparameter settings.

Minor

Features:
1. Added option for ``permutation='learnable'`` when creating an ``InvertibleNetwork``
2. Added option for ``coupling_design in ["affine", "spline", "interleaved"]`` when creating an ``InvertibleNetwork``
3. Simplified passing additional settings to the internal networks. For instance, you
can now simply do
``inference_network = InvertibleNetwork(num_params=20, coupling_net_settings={'mc_dropout': True})``
to get a Bayesian neural network.