Ax

Compatibility: Requires botorch >= 0.3.3

Major changes

• Support for new Botorch features, including Contextual Bayes Opt, Input Warping, etc. See the latest Botorch release for more information: https://github.com/pytorch/botorch/releases/tag/v0.3.3
• Improvements to storage-related performance by aligning SQLAlchemy and Ax classes better (e.g. #422)

• Support and new tutorial for Multi-Objective Optimization (MOO)

Fixed re-release of v0.1.15 and v0.1.16 that were deployed incorrectly.

Minor version bump to resolve some bugs with the last release (#390, #391, #387)

Modular Botorch

• Introduces a new set of abstractions for Botorch models to support customizable, storage-friendly botorch components (e.g. Surrogate, Acquisition).
• See #374 for more information
• A draft tutorial is available at: https://github.com/facebook/Ax/blob/master/tutorials/botorch_modular.ipynb

Logger redesign

• Our logger APIs have been refactored to use a new naming scheme (module name, via getLogger(__name__)) and use a different handler/propagation scheme (1 root handler instead of N+1 child handlers). These changes play nicely with core features of Python's logging API and enable use cases such as #316.

MultiObjective API Improvements

• Introduces some new MOO modeling APIs such as MultiObjectiveTorchModelBridge to support our growing list of MOO features. See 272b6dd

Misc Fixes
#351, #382

• Minor bug fixes (storage of function kwargs to BoTorch models, etc.),
• Additional documentation on the website.

Minor release, including:

• Some bug fixes and code clean up.
• Additional metadata from candidate generation exposed via GeneratorRun

Misc. bug fixes and improvements

• New experiment method add_tracking_metrics.
• Speed up experiment save/load.
• Fix unexpected or buggy MultiObjective behavior.
• Fixed bug that caused forced AxClient to require SQLAlchemy
• Add experiment method trials_by_status.
• Expand the Service API tutorial to cover handling common exceptions and to discuss how to manage parallelism in Service API optimization.
• Fix Raytune and Service API tutorials.
• Allow for more detailed parallelism configuration in choose_generation_strategy dispatch utility.

• Generation strategy (and consequently the Service API aka AxClient) now enforces maximum parallelism per step.
• Parallelism limitation will mostly be relevant for the Bayesian optimization phase of an Ax experiment; if you run into a MaxParallelismReachedException, which will mean that more trials need to be completed with data before generating new ones.
• Sequential optimization with low parallelism is key to good Bayesian optimization performance (achieving optimal results in a low number of trials), but if computational resources are not a constraint, maximum parallelism for a given generation step can be altered by providing a custom generation strategy to AxClient
• There is also now a DataRequiredError raised by generation strategy (and thus appearing in Service API).
• This error is meant to designate when more trials need to be completed with data before Ax can transition to the next model in the generation strategy for a given optimization."

Core

• Introduces new methods complete_trial() and update_trial(), which allow for full or partial completion of trials, respectively
• Introduces a DataRequiredError error, allowing generation strategies to signal when more data is necessary before new suggestions can be generated
• Adds support for custom generation strategies as a part of the managed loop
• Some storage API cleanup to deprecate the DISPATCHED status

Modelling/Botorch

• Bumps up the botorch version to 0.2.1, the latest as of this release
• Introduces new linear embedding methods (ALEBO, REMBO, HeSBO)
• Adds functionality for benchmarking external methods that implement a minimization interface like scipy.optimize.minimize.

Tutorials

• Fixes to the multi-task tutorial
• Fixes to the Service API tutorial

• Enable de-duplication by default for Sobol Generator
• Improved handling of out-of-design points in Modelbridge
• Use Sobol for small integer spaces in GenerationStrategy (ac91f1d)
• New RayTune tutorial
• Multiple tutorial fixes
• Expose best_point in BotorchModel
• Plotly 4.0 compatibility fixes
• BoTorch 0.2 compatibility

• Bugfix with use of SQL backend for generation strategy saving,
• Improved Service API visualization and tutorial,
• Plotting in JupyterLabs,
• Miscellaneous bugfixes and improvements.

Miscellaneous fixes and improvements, including:

• SQA storage and JSON serialization of GenerationStrategy;
• Can now interrupt and resume a Service API optimization (optimization using AxClient);
• Optimization is no longer assumed to be noiseless by default in Service and Loop API optimizations;
• Service API trial generation bug is fixed, etc.

NOTE: This release depends on a bleeding edge version of BoTorch, and should not be used. Please use 0.1.5.

Misc. bug fixes and improvements

Small bug fixes + documentation updates
Bug fixes for issues: #76, #81, #83, #89, #91, #93

v0.1.1 Release of Ax, corresponding to PyPI release