Randomly Projected Additive GPs git repo

This repo contains implementations and experiment code for the paper Randomly Projected Additive Gaussian Processes for Regression

Requirements

• Python > 3.0
• GPyTorch >= 1.0
• PyKeOps >= 1.2

Files

• config_template.py: Template configuration file for dataset file locations, etc. Rename to config.py and replace with your file configurations. UCI datasets referenced in the experiments may be downloaded here.
• gp_experiment_runner.py: Command-line endpoint used for running batches of experiments.
• synthetic_test_script.py: A simple script for running synthetic experiments.
• rp.py: Generating (random) projection matrices, including a routine for generating diversified projection matrices (useed in DPA-GP).
• training_routines.py: A collection of routines used to construct, train, and test GPs in this project.
• test.py: a suite of unit tests.
• utils.py: Utilities that are reused and don't live in a particular section of the project.

Subpackages

• gp_models: Encapsulates the model (and kernel) definitions for kernels and models used.
• fitting: Encapsulates methods for learning. Currently, only optimization-based methods are available, as opposed to, e.g., sampling.

Folders

• model_specs: Model specification .json files. These are used to store and re-use the configuration of models.
• run_scripts: Re-used/example command-line calls to gp_experiment_runner.py.

UCI Data Sets

To download the UCI data sets used for benchmarks, download them from Andrew Gordon Wilson's home page. See config_template.py for details on how these files are expected to be organized in accordance with your configurations.