v0.1.14

What's Changed

• Add a json encoder for metadata by @copybara-service in #1009
• Add failing designer for testing. by @copybara-service in #1010
• Allow for linear kernel for ARD in Python GP. by @copybara-service in #1012
• Fix early stopping colab by @copybara-service in #1014
• Modernize @jaxtyped code by @copybara-service in #1015
• Support reverse log in keras embedder and small edge case fix in core converter. by @copybara-service in #1017
• Support SimpleKd experimenter returning relative objective value. by @copybara-service in #1011
• Propagate selected keys from global namespace to metadata namespace. by @copybara-service in #1023
• Simplify some SQL datastore code by @copybara-service in #1022
• gp_bandit designer rejects empty search spaces. by @copybara-service in #1024
• No need for union since vz.Trial is already a subclass by @copybara-service in #1025
• Debug shifting experimenter due to OOB values (and revert to previous code). by @copybara-service in #1019
• Apply suggestion from ConnorBaker in #1029 by @copybara-service in #1033
• Remove Python 3.7 support and fix ugly formatting in pyglove core by @copybara-service in #1035
• Allow for mutation rate schedules for NSGA by @copybara-service in #1034
• Update benchmark result structure to incorporate spec gen name. by @copybara-service in #1020
• Adds a new error type (not yet fully supported) that allows Pythia algorithms by @copybara-service in #1038
• Some support for PythiaFallbackError. by @copybara-service in #1039
• Adds benchmarks testing. by @copybara-service in #1040
• Add NormalizedSimpleRegert and WinRateSimpleRegret convergence curve comparators. by @copybara-service in #1041
• Creating scheduled GP-Bandit designer. by @copybara-service in #1043

Full Changelog: v0.1.13...v0.1.14

v0.1.13

What's Changed

• Remove logging errors when the convertor rounds up values, which is not an uncommon occurrence (e.g. RandomDesigner, EagleStrategyDesigner). by @copybara-service in #950
• Input validation for add_categorical_param. by @copybara-service in #952
• Fix the type error. nn.module.get_variable returns a regular dict now. by @copybara-service in #957
• Fixes Eagle optimizer trials padding by @copybara-service in #960
• ParameterEntropyHparamScorerFactory for scoring exploration behavior by @copybara-service in #970
• Add prior study setters into benchmark runners. by @copybara-service in #984
• Update documentation: Users are now allowed to write to namespaces other than by @copybara-service in #988
• Introduce stateful curve converters. by @copybara-service in #987
• Add support for conditionals in TrialToArrayConverter. by @copybara-service in #1001
• Minor edits by @copybara-service in #1000

Full Changelog: v0.1.12...v0.1.13

v0.1.12

What's Changed

• Small aesthetic update to Ray benchmark RTD by @copybara-service in #914
• Add a lot of tracing trackers. by @copybara-service in #913
• Fix broken pytest_ray installation by @copybara-service in #918
• Add multi-objective acquisitions and integrate them into GPBandit. by @copybara-service in #909
• Add profiling around seed trial generation to help debug slow runs. by @copybara-service in #925
• Meta learning designer: 1) multi-metric trial support, 2) updated default schedules, 3) additional testing and documentation. by @copybara-service in #933
• Add counter metrics with metadata. by @copybara-service in #931
• Use QuasiRandomDesigner in EagleStrategy. by @copybara-service in #943
• Clean up & bugfix for random rotation matrix generation. by @copybara-service in #941
• Remove redundant logging from Eagle designer for expediting benchmarking jobs / convergence tests. by @copybara-service in #946

Full Changelog: v0.1.11...v0.1.12

v0.1.11

What's Changed

• Expose PolicyFactory into Pythia import. by @copybara-service in #905
• General Meta-Learning Framework by @copybara-service in #877
• Debug ensemble_designer to handle with Active Trials. by @copybara-service in #910

Full Changelog: v0.1.10...v0.1.11

v0.1.10

What's Changed

• Clarify that in multi-worker situations, workers are expected to provide client_id in the suggest() call, and not during client construction. by @copybara-service in #875
• Expose DesignerPolicy in algorithms init file by @copybara-service in #894
• Update Vizier version to 0.1.10 by @copybara-service in #896

Full Changelog: v0.1.9...v0.1.10

v0.1.9

What's Changed

• Add PlotElements and BenchmarkRecord for easy plotting. by @copybara-service in #878

Full Changelog: v0.1.8...v0.1.9

v0.1.8

What's Changed

• 'all' pip README update by @copybara-service in #851

Full Changelog: v0.1.7...v0.1.8

v0.1.7

What's Changed

• Predict Kumamon individual metrics. by @copybara-service in #739
• Add GP Bandit tests for VizierSearcher by @copybara-service in #750
• Make the number of unpadded features a dynamic arg for acquisition function optimization so that padding/masking the feature dimension avoids retracing. by @copybara-service in #757
• Add a Converter from trials to ContinuousAndCategoricalArray. by @copybara-service in #754
• Minor cleanups by @copybara-service in #788
• Use Categorical kernel in GP bandit. by @copybara-service in #784
• Convert vectorized optimizers to operate on continuous and categorical data. by @copybara-service in #793
• Follow up to the VectorizedBase refactor with a couple fixes: by @copybara-service in #797
• Add arch gym to highlights by @copybara-service in #804
• Add the wide LCB and delayed trust region application options to the acquisition-based trust region. by @copybara-service in #805
• Silence some pytype errors. by @copybara-service in #820
• Minor updates to pythia docs, string reps, and TODOs by @copybara-service in #821
• Internal change. by @copybara-service in #829
• Advance attrs version to 23.1.0 by @copybara-service in #830
• Add batching/ensembling support for mean_fn. by @copybara-service in #834

Full Changelog: v0.1.6...v0.1.7

v0.1.6

What's Changed

• Fix qUCB test following a change to TFP. Rename qUCB parameter "coefficient" for consistency with UCB. by @copybara-service in #691
• Bug fix/add missing test cases for Optax ARD optimizer. by @copybara-service in #692
• Add option to pad number of observations and dimensions of inputs. by @copybara-service in #627
• Pyvizier needs to include one more definition. by @copybara-service in #693
• Update Adaptive Ensembling to use observation probs. by @copybara-service in #695
• Clarify jit_loop arg in vectorized_base.py by @copybara-service in #705
• Fix the flag error in runtime context. #698 by @copybara-service in #700
• Update Basic Vizier documentation by @copybara-service in #716
• Update all Jax requirements to avoid breaking Github workflows by @copybara-service in #722
• Disable benchmark tests for now due to Dopamine failures. by @copybara-service in #730
• Remove 3.x from github workflow by @copybara-service in #731

Full Changelog: v0.1.5...v0.1.6

v0.1.5

What's Changed

• Create input warping based on Kumaraswamy CDF. by @copybara-service in #562
• Fix Numpy dtype error in acquisitions_test. by @copybara-service in #605
• Add a test to the RandomDesigner to confirm that it respects LOG scaling. by @copybara-service in #596
• Fix Eagle slowness for large studies by @copybara-service in #609
• Add typing to output warping by @copybara-service in #613
• Use time as seed by @copybara-service in #614
• Fix pytype failures related to teaching pytype about NumPy scalar types. by @copybara-service in #604
• Add RayTune SearchSpace and Experimenter Converters and test on mock study. by @copybara-service in #615
• Update GoogleDesignerFactory to use a designer factory. by @copybara-service in #611
• Remove GP hparam ensemble since ensemble members are usually the same with the LBFGSB optimizer, and to reduce latency associated with jax.vmap with the forthcoming JAX Eagle optimizer. Change best_n=1 behavior to squeeze out the singleton dimension of the GP hparams and not use vmap, since unnecessary vmap is a speed footgun. by @copybara-service in #618
• Normalize loss value by the number of observations only when using SGD-based GP hparam optimizers (so a consistent learning rate can be used regardless of the number of observations). Previously, this was baked into the loss function definition; now that we're using L-BFGS-B by default, this is unnecessary and possibly confusing. by @copybara-service in #602
• Update Pytypes after recent changes to typing in JAX/Numpy, and following our decision to use more specific types (instead of chex.Array) where possible. by @copybara-service in #624
• Fix a bug where Constraint objects returned by get_constraints did not account for mean_fn parameters. by @copybara-service in #633
• Added utils/performance_utils.py with a decorator that tracks function runtime duration. by @copybara-service in #638
• Make 'num_metric' the last dimension in LinearOutputWarper by @copybara-service in #648
• Kumamon: apply linear output warping and compute acquisition for separable kernel using sampling. by @copybara-service in #649
• Add more logs to DesignerPolicy to better track performance. by @copybara-service in #653
• Update mapping logic for Raytune. by @copybara-service in #657
• Add L-BFGS-B Acquisition function optimizer. by @copybara-service in #658
• Move NestedDictRAMDatastore to a new file and move metadata operations into metadata_util.py. by @copybara-service in #673
• Update Predictor pydoc. by @copybara-service in #676
• Update predict RTD colab. by @copybara-service in #681
• Fixes Trial default creation_time by @copybara-service in #684

Full Changelog: v0.1.4...v0.1.5