Generalised-Gaussian-Processes

Fully Bayesian Inference in GPs - learning hyperparameter distributions with different likelihoods and inference techniques. This repo is a work in progress.

Likelihoods

1. Gaussian (Regression)
2. Bernoulli Probit (Classification)
3. SoftMax (Multi-Class)
4. Log Cox Posisson (Regression)

Models

1. SGPR (Titsias, 2009)
2. SVGP (Hensman, 2013 / Hensman, 2015)
3. BayesianSVGP (new)
4. BayesianSGPR_HMC (new)

Inference methods

1. Hamiltonian Monte Carlo (pymc3)
2. Stochastic Variational Inference (SVI) (gpytorch/pytorch)

Future work

3. Elliptical Slice Sampling (pymc3)
4. Dynamic Nested Sampling (dynesty)

Code Layout

Please set the working directory to the parent folder which containst the following sub-folders

utils/ - Data loading and visualisation utilities.

experiments/ - Scripts for running experiments and generating plots.

models/ - Classes and methods which encapsulate inference.

results/ - Tables, logging and directory for plots.