Skip to content

google-research/hyperbo

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

60 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

HyperBO - Prior Discovery

A Jax/Flax codebase for the algorithm in HyperBO described in Pre-trained Gaussian processes for Bayesian optimization.

Research paper | Blog post | Colab Notebook | PD1 benchmark

Disclaimer: This is not an officially supported Google product.

Tutorial

Follow HyperBO's Colab Notebook or Jupyter Notebook.

Also see tests for a more comprehensive understanding of the usage.

Installation

We recommend using Python 3.7 or 3.9 for stability.

To install the latest development version inside a virtual environment, run

python3 -m venv env-pd
source env-pd/bin/activate
pip install --upgrade pip
pip install "git+https://github.com/google-research/hyperbo.git#egg=hyperbo"

PD1 benchmark

PD1 is a new hyperparameter tuning benchmark for optimizing deep learning models. To download the PD1 dataset, please copy and paste the following link to your browser's address bar.

http://storage.googleapis.com/gresearch/pint/pd1.tar.gz

See pd1/README.txt for more information. The data is licensed under the CC-BY 4.0 license.

If you'd like to use the evaluations at each training step, the relevant columns of the data frame are

'valid/ce_loss'
'train/ce_loss',
'train/error_rate',

etc. They will hold arrays aligned with the global_step column that indicates what training step the measurement was taken at.

See the "best_*" columns for the best measurement achieved over training.

GPax

GPax is a modular implementation of Gaussian processes used by HyperBO based on Tensorflow Probability with Jax backend.

Citation

Please cite our work if you would like to use the code.

@article{wang2023hyperbo,
  title={{Pre-trained Gaussian processes for Bayesian optimization}},
  author={Zi Wang and
          George E. Dahl and
          Kevin Swersky and
          Chansoo Lee and
          Zachary Nado and
          Justin Gilmer and
          Jasper Snoek and
          Zoubin Ghahramani},
  journal={arXiv preprint arXiv:2109.08215},
  year={2023}
}