Open Source Vizier: Reliable and Flexible Black-Box Optimization.

Google AI Blog | Getting Started | Documentation | Installation | Citing Vizier

What is Open Source (OSS) Vizier?

OSS Vizier is a Python-based service for black-box optimization and research, based on Google Vizier, one of the first hyperparameter tuning services designed to work at scale.

OSS Vizier's distributed client-server system. Animation by Tom Small.

Getting Started

As a basic example for users, below shows how to tune a simple objective using all flat search space types:

from vizier.service import clients
from vizier.service import pyvizier as vz

# Objective function to maximize.
def evaluate(w: float, x: int, y: float, z: str) -> float:
  return w**2 - y**2 + x * ord(z)

# Algorithm, search space, and metrics.
study_config = vz.StudyConfig(algorithm='GAUSSIAN_PROCESS_BANDIT')
study_config.search_space.root.add_float_param('w', 0.0, 5.0)
study_config.search_space.root.add_int_param('x', -2, 2)
study_config.search_space.root.add_discrete_param('y', [0.3, 7.2])
study_config.search_space.root.add_categorical_param('z', ['a', 'g', 'k'])
study_config.metric_information.append(vz.MetricInformation('metric_name', goal=vz.ObjectiveMetricGoal.MAXIMIZE))

# Setup client and begin optimization. Vizier Service will be implicitly created.
study = clients.Study.from_study_config(study_config, owner='my_name', study_id='example')
for i in range(10):
  suggestions = study.suggest(count=1)
  for suggestion in suggestions:
    params = suggestion.parameters
    objective = evaluate(params['w'], params['x'], params['y'], params['z'])
    suggestion.complete(vz.Measurement({'metric_name': objective}))

Documentation

OSS Vizier's interface consists of three main APIs:

• User API: Allows a user to setup an OSS Vizier Server, which can host black-box optimization algorithms to serve multiple clients simultaneously in a fault-tolerant manner to tune their objective functions.
• Developer API: Defines abstractions and utilities for implementing new optimization algorithms for research and to be hosted in the service.
• Benchmarking API: A wide collection of objective functions and methods to benchmark and compare algorithms.

Additionally, it contains advanced API for:

• Tensorflow Probability: For writing Bayesian Optimization algorithms using Tensorflow Probability and Flax.
• PyGlove: For large-scale evolutionary experimentation and program search using OSS Vizier as a distributed backend.

Please see OSS Vizier's ReadTheDocs documentation for detailed information.

Installation

Most common: To tune objectives using our default state-of-the-art JAX-based Bayesian Optimizer, run:

pip install google-vizier[jax]

To install a minimal version that consists of only the core service and client API from requirements.txt, run:

pip install google-vizier

For full installation to support all algorithms and benchmarks, run:

pip install google-vizier[extra]

For specific installations, you can run:

pip install google-vizier[X]

which will install additional packages from requirements-X.txt, such as:

• requirements-jax.txt: Jax libraries shared by both algorithms and benchmarks.
• requirements-tf.txt: Tensorflow libraries used by benchmarks.
• requirements-algorithms.txt: Additional repositories (e.g. EvoJAX) for algorithms.
• requirements-benchmarks.txt: Additional repositories (e.g. NASBENCH-201) for benchmarks.
• requirements-test.txt: Libraries needed for testing code.

Check if all unit tests work by running run_tests.sh after a full installation. OSS Vizier requires Python 3.10+, while client-only packages require Python 3.7+.

Citing Vizier

If you found this code useful, please consider citing the OSS Vizier paper as well as the Google Vizier paper. Thanks!

@inproceedings{oss_vizier,
  author    = {Xingyou Song and
               Sagi Perel and
               Chansoo Lee and
               Greg Kochanski and
               Daniel Golovin},
  title     = {Open Source Vizier: Distributed Infrastructure and API for Reliable and Flexible Black-box Optimization},
  booktitle = {Automated Machine Learning Conference, Systems Track (AutoML-Conf Systems)},
  year      = {2022},
}
@inproceedings{google_vizier,
  author    = {Daniel Golovin and
               Benjamin Solnik and
               Subhodeep Moitra and
               Greg Kochanski and
               John Karro and
               D. Sculley},
  title     = {Google Vizier: {A} Service for Black-Box Optimization},
  booktitle = {Proceedings of the 23rd {ACM} {SIGKDD} International Conference on
               Knowledge Discovery and Data Mining, Halifax, NS, Canada, August 13
               - 17, 2017},
  pages     = {1487--1495},
  publisher = {{ACM}},
  year      = {2017},
  url       = {https://doi.org/10.1145/3097983.3098043},
  doi       = {10.1145/3097983.3098043},
}