Guepard aims at building faster Gaussian Process (GP) models by constructing and aggregating sub-models based on subsets of the data. It is based on GPflow and implements various aggregation methods for GP ensembles:
- Equivalent Observation as described in the AISTATS submission
- Nested GPs [Rullière 2018]
- Barycenter GP [Cohen 2020]
- Several classic baselines: (generalised) Product of Expert, (robust) Bayesian committee machine, etc
To install the library run
poetry install
in a terminal at the root of the repo
The project uses black, isort, and flake8 for code formating and linting
poetry run task format
and it uses pytest for testing
poetry run task check
poetry run task test
- Didier Rullière, Nicolas Durrande, François Bachoc, and Clément Chevalier. Nested Kriging predictions for datasets with a large number of observations. Statistics and Computing, 2018.
- Samuel Cohen, Rendani Mbuvha, Tshilidzi Marwala, and Marc Peter Deisenroth. Healing products of Gaussian process experts. ICML 2020.