Meta-BO based on PACOH / F-PACOH

The repository implements the UCB Bayesian optimization algorithm with support for the following models

• Vanilla GP
• PACOH-GP-MAP (https://arxiv.org/abs/2002.05551)
• PACOH-GP-SVGD (https://arxiv.org/abs/2002.05551)
• F-PACOH-GP (https://arxiv.org/abs/2106.03195)

Installation

To install the minimal dependencies needed to use the meta-learning algorithms, run in the main directory of this repository

pip install -e .

Note that if you want jax with gpu support and don't have it installed already, you can additionally run the command

pip install --upgrade jax[cuda] -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

See also the jax documentation for that.

Usage

First, add the root of this repository to PYTHONPATH, e.g. using sys.path.append(<path_to_repo>) within your virtual environment. Then, simply use from pacoh.models import ... to import the model you want to meta-learn with.

Check out the models directory for more documentation of how to use the models and the parameters available.

Basic Usage

First, add the root of this repository to PYTHONPATH. The following code snippet provides a basic usage example of F-PACOH:

from pacoh.bo.meta_environment import RandomMixtureMetaEnv
from pacoh.models import F_PACOH_MAP_GP
import numpy as np

# generate some meta train and meta test data
meta_env = RandomMixtureMetaEnv()
num_train_tasks = 20
meta_train_data = meta_env.generate_uniform_meta_train_data(
    num_tasks=num_train_tasks, num_points_per_task=10
)
x_train, y_train, _, __ = meta_env.generate_uniform_meta_valid_data(
    num_tasks=1, num_points_context=10, num_points_test=10
)[0]
x_pred = np.linspace(meta_env.domain.l, meta_env.domain.u, num=100)


# initialize F-PACOH model
gp_model = F_PACOH_MAP_GP(input_dim=1, output_dim=1, domain=meta_env.domain, num_tasks=num_train_tasks, 
                          weight_decay=1e-4, prior_factor=1e-2, task_batch_size=2, lr=1e-3
)

# meta train the model
gp_model.meta_fit(
    meta_train_data, num_iter_fit=200
)

# meta predict
posterior_mean, posterior_std = gp_model.meta_predict(x_train, y_train, x_pred, return_density=False)

Citing

For usage of the algorithms provided in this repo for your research, we kindly ask you to acknowledge the two papers that formally introduce them:

@InProceedings{rothfuss21pacoh,
  title = 	 {PACOH: Bayes-Optimal Meta-Learning with PAC-Guarantees},
  author =       {Rothfuss, Jonas and Fortuin, Vincent and Josifoski, Martin and Krause, Andreas},
  booktitle = 	 {Proceedings of the 38th International Conference on Machine Learning},
  pages = 	 {9116--9126},
  year = 	 {2021}
 }

@InProceedings{rothfuss2021fpacoh,
  title={Meta-learning reliable priors in the function space},
  author={Rothfuss, Jonas and Heyn, Dominique and Chen, Jinfan and Krause, Andreas},
  booktitle={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}