Release v4.4.0

Improvements/fixes

TF datasets are now passed to DeepEnsemble.model.fit. This change enables support for training over a specified number of steps (rather than a set number of epochs) for small datasets, which may require data repetition to achieve the desired step count. #907

An optional predict_fn parameter has been added to IndependentReparametrizationSampler.init. This parameter allows generating samples where the mean and variance come from sources other than model.predict. This feature is particularly useful for drawing samples from models that separate epistemic and aleatoric uncertainty, providing greater flexibility in controlling the sources of uncertainty in the generated samples. #903

Release v4.3.0

Improvements/fixes
implement predict_y and predict_noise for Deep Ensemble models, and update predict to more appropriately return the epistemic uncertainty #894
fix jitter bug in Deep Ensemble trajectory sampler #900

Full Changelog: v4.2.4...v4.3.0

Release v4.2.4

Fixes
Remove default independent reparametrization sampler jitter (but ensure positive variance) (#888)
Test against TF 2.14 (#895)

Full Changelog: v4.2.3...v4.2.4

Release 4.2.3

Fixes
Fix DeepEnsemble model serialization for some versions of tensorflow (#892)

Full Changelog: v4.2.2...v4.2.3

Release 4.2.2

Fixes

• Don't use the same seed when sampling from the subspaces of a product space [#885]
• Fix failing ray tutorial [#856]

Full Changelog: v4.2.1...v4.2.2

Release 4.2.1

Fixes
Make one hot encoder compilable (#880)
Improve dataset_len error message (#879)

Full Changelog: v4.2.0...v4.2.1

Release 4.2.0

Improvements/fixes
Avoid duplicating initial points when using vectorization (#875)

Full Changelog: v4.1.0...v4.2.0

Release 4.1.0

Improvements/fixes
Query point encoders for Deep GP models (#873)

Full Changelog: v4.0.1...v4.1.0

Release 4.0.1

Improvements/fixes
Don't create additional feature when one-hot-encoding two-value categories (#869)
Support and test against Tensorflow 2.16 (#858)

Note that (like tensorflow-probability and GPflow) Trieste uses Keras 2. Since TF 2.16 defaults to using Keras 3, tf.keras (and Keras optimizers such as Adam) must now be imported from the tf_keras package instead. Alternatively, you can import tf_keras from the gpflow.keras module, which will automatically select the right source depending on which version of TF is installed.

Full Changelog: v4.0.0...v4.0.1

v4.0.0

Breaking changes
This release includes a minor breaking change:

As part of #864, a number of built in model classes such as GaussianProcessRegression and DeepEnsembleModel have been updated to support optional query point encoders. This involved moving the implementation of public methods such as predict to new internal methods called predict_encoded etc that work on the encoded query points. Any user-defined class that overrode the public methods should therefore switch to overriding the *_encoded internal methods instead.

New features
Query point encoders for models (#864)

Improvements/fixes
Categorical trust regions (#865)

Full Changelog: v3.4.0...v4.0.0