gp_upper contains some extra functionality for computing upper bounds to Gaussian process regression marginal
likelihoods, that isn't included in GPflow. Currently, this is minimisation of the
upper bound, to tighten it. Usage can be deduced from the notebook, which also discusses how the upper bound can be used
to diagnose over-estimation the marginal likelihood by FITC.
The package is pretty tiny, but I added setup.py just in case. Install using python setup.py develop.
A unit test is included for easy verification of correct functioning.
nosetests testing --nologcapture --with-coverage --cover-package=gp_upper --cover-erase
- Notebook which shows the effect of optimising Z after training with FITC and VFE models.
- Add option for using different bounds on maximum eigenvalue.
This project was made to add some extra functionality not absorbed into the GPflow core. While I'll try to keep it up-to-date, I'm not giving it the same guarantees as GPflow. If something is wrong, or something breaks due to an update in TensorFlow/GPflow/whatever, feel free to raise an issue or submit a PR.