Orthogonal sparse variational Gaussian processes (SVGP's) in TensorFlow.
Orthogonal methods [1,2] extend standard SVGP approaches [3] by splitting the inducing points learned by the model into two orthogonal sets with structured covariance. The resulting approximate posterior can be decomposed into several independent (orthogonal) Gaussian processes and both training and prediction involves reduced complexity matrix operations as compared with standard SVGP. This project implements an orthogonal SVGP framework using GPflow/TensorFlow and explores whether orthogonal methods are indeed able to provide high quality posterior approximations with fewer inducing points than standard SVGP methods and/or at reduced computational cost.
This project relies heavily on GPflow, as well as NumPy, Matplotlib, Jupyter,
TensorFlow and TensorFlow Probability. Note that TensorFlow Probability
releases are tightly coupled to TensorFlow; please ensure that the TensorFlow
and TensorFlow Probability versions you have installed are compatible. You may
wish to manually install an older version of TensorFlow Probability to address
compatibility issues. Properly installing TensorFlow depends on your operating
system and hardware, but a sample environment.yml file containing complete
project dependencies may be found in the env/ directory which can be used
to set up the environment with Anaconda for M1 Mac users.
$ cd env/
$ conda env create -f environment.ymlAll source code was written by Oskar Fernlund at Imperial College London in support of his 2022 Master's thesis. Implementation of the orthogonal framework depends heavily on GPflow and roughly follows its model API. For a complete list of GPflow contributors, see GPflow's GitHub page. Many thanks to Mark van der Wilk and Anish Dhir for all their insightful feedback and support.
Use and distribution of the source code is covered by the Apache-2.0 license.