GPflow is a package for building Gaussian process models in python, using TensorFlow. It was originally created and is now managed by James Hensman and Alexander G. de G. Matthews. The full list of contributors (in alphabetical order) is Alexis Boukouvalas, Ivo Couckuyt, Keisuke Fujii, Zoubin Ghahramani, David J. Harris, James Hensman, Pablo Leon-Villagra, Daniel Marthaler, Alexander G. de G. Matthews, Tom Nickson, Valentine Svensson and Mark van der Wilk. GPflow is an open source project so if you feel you have some relevant skills and are interested in contributing then please do contact us.
Please see instructions on the main TensorFlow webpage. You will need version 0.11. We find that for most users pip installation is the fastest way to get going.
GPflow includes some tensorflow extensions that are compiled when you run setup.py. For those interested in modifying the source of GPflow, we recommend
python setup.py develop
but installation should work well too:
python setup.py install
You can run the tests with python setup.py test.
GPflow has origins in GPy by the GPy contributors, and much of the interface is intentionally similar for continuity (though some parts of the interface may diverge in future). GPflow has a rather different remit from GPy though:
- GPflow leverages TensorFlow for faster/bigger computation
- GPflow has much less code than GPy, mostly because all gradient computation is handled by tensorflow.
- GPflow focusses on variational inference and MCMC -- there is no expectation propagation or Laplace approximation.
- GPflow does not have any plotting functionality.
GPflow has a slew of kernels that can be combined in a similar way to GPy (see this tutorial). As for inference, the options are currently:
For GP regression with Gaussian noise, it's possible to marginalize the function values exactly: you'll find this in GPflow.gpr.GPR. You can do maximum likelihood or MCMC for the covariance function parameters (notebook).
It's also possible to do Sparse GP regression using the GPflow.sgpr.SGPR class. This is based on [4].
For non-Gaussian likelihoods, GPflow has a model that can jointly sample over the function values and the covariance parameters: GPflow.gpmc.GPMC. There's also a sparse equivalent in GPflow.sgpmc.SGPMC, based on a recent paper [1]. This notebook introduces the interface.
It's often sufficient to approximate the function values as a Gaussian, for which we follow [2] in GPflow.vgp.VGP. In addition, there is a sparse version based on [3] in GPflow.svgp.SVGP. In the Gaussian likelihood case some of the optimization may be done analytically as discussed in [4] and implemented in GPflow.sgpr.SGPR . All of the sparse methods in GPflow are solidified in [5].
The following table summarizes the model options in GPflow.
| Gaussian likelihood |
Non-Gaussian (variational) |
Non-Gaussian (MCMC) |
|
|---|---|---|---|
| Full-covariance | GPflow.gpr.GPR |
GPflow.vgp.VGP |
GPflow.gpmc.GPMC |
| Sparse approximation | GPflow.sgpr.SGPR |
GPflow.svgp.SVGP |
GPflow.sgpmc.SGPMC |
[1] MCMC for Variationally Sparse Gaussian Processes J Hensman, A G de G Matthews, M Filippone, Z Ghahramani Advances in Neural Information Processing Systems, 1639-1647
[2] The variational Gaussian approximation revisited M Opper, C Archambeau Neural computation 21 (3), 786-792
[3] Scalable Variational Gaussian Process Classification J Hensman, A G de G Matthews, Z Ghahramani Proceedings of AISTATS 18, 2015
[4] Variational Learning of Inducing Variables in Sparse Gaussian Processes. M Titsias Proceedings of AISTATS 12, 2009
[5] On Sparse variational methods and the Kullback-Leibler divergence between stochastic processes A G de G Matthews, J Hensman, R E Turner, Z Ghahramani Proceedings of AISTATS 19, 2016
James Hensman was supported by an MRC fellowship and Alexander G. de G. Matthews was supported by EPSRC grant EP/I036575/1.