Implements a variety of Gaussian process models exploiting the "structured"
assumption that one has inputs that are formed as a Cartesian product as well
as a kernel that is separable so that one may decompose the associated kernel
matrices as Kronecker products for a representation that is computationally
efficient in terms of both time and memory.
The models are built on top of GPflow, and
the computational backend is TensorFlow.
First, install GPflow. (Note: this repo is designed to work with this fork.)
Then, simply python setup.py install as usual.
- SGPR: Structured GP for regression
- SGPLVM: Structured Bayesian Gaussian process latent variable model
- SWGP: Structured Bayesian warped Gaussian processes
See [Atkinson and Zabaras, 2018] for more information.
Contact Steven Atkinson or Nicholas Zabaras with questions or comments.