Skip to content

Latest commit

 

History

History
85 lines (63 loc) · 3.95 KB

CHANGELOG.rst

File metadata and controls

85 lines (63 loc) · 3.95 KB
Raw

Change log

1.0.2b Series

Minor

  1. Added option for residual networks in coupling blocks using a residual=True in the coupling layer config.

1.0.3b Series

Major (Breaking)

  1. Coupling layers were refactored to ensure consistency. Old checkpoints may no longer load.

Minor (Features)

  1. Added attention.py module containing helper networks for building transformers
  2. Added SetTransformer class in summary_networks.py as a viable alternative to DeepSet summary networks.
  3. Added TimeSeriesTransformer class in summary_networks.py as a viable alternative to SequentialNetworks summary networks.
  4. Added plot_z_score_contraction() diagnostic in diagnostics.py for gauging global inferential adequacy
  5. Added Orthogonal in helper_networks.py for learnable generalized permutations.

1.1 Series

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. 4. PMPNetwork has been added for model comparison according to findings in https://arxiv.org/abs/2301.11873 5. HierarchicalNetwork wrapper has been added to act as a summary network for hierarchical Bayesian models according to https://arxiv.org/abs/2301.11873 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

1.1.2 Series

  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

1.1.3 Series

  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

1.1.4 Series

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.