This repository is the official implementation of the methods in the publication:
- O. Hamelijnck, W.J. Wilkinson, N.A. Loppi, A. Solin, and T. Damoulas (2021). Spatio-temporal variational Gaussian processes. In Neural Information Processing Systems (NeurIPS). [arXiv]
@inproceedings{hamelijnck2021spatio,
title={Spatio-Temporal Variational {G}aussian Processes},
author={Hamelijnck, Oliver and Wilkinson, William and Loppi, Niki and Solin, Arno and Damoulas, Theodoros},
booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
year={2021},
}
This has been tested on a Macbook Pro. All spatio-temporal VGP models have been implemented within the Bayes-Newton package.
We recommend using conda:
conda create -n spatio_gp python=3.7
conda activate spatio_gp
Then install the required python packages:
pip install -r requirements.txt
All data, preprocessed and split into train-test splits used in the paper is provided at https://doi.org/10.5281/zenodo.4531304. Download the folder and place the corresponding datasets into experiments/*/data
folders.
We also provide scripts to generate the data manually:
make data
which will download the relevant London air quality and NYC data, clean them, and split into train-test splits.
To run all experiments across all training folds run:
make experiments
To run an individual experiment refer to the Makefile
.
GPFlow2
: https://github.com/GPflow/GPflowGPYTorch
: https://github.com/cornellius-gp/gpytorch
This software is provided under the MIT license.