This document details some of the progress made on the library
- Add tests from DSDGP
- Add method to initialize model params
- Test DSDGP on toy example (add to docs)
- Test DSDGP and its optins further
- Work on improving documentation
- Look into multitask learning extensions
- Sequential can be intialised with a partial list of layers, e.g.
[InputLayer, HiddenLayer, HiddenLayer]
. Should we disallow this in favour of the correct layer structure, i.e.OutputLayer
at the end? - Currently
Sequential
only works by modifyingGPflow
source code. If pull request accepted nothing to be done here. Else, design needs to be changed. InputLayer
takes an initial value for Z, maybe better to disallow this and add it to the model object insteadSequential
has an_initialized
private attribute. Turn this into property.- So far, multikernel layers have a set of inducing inputs for each kernel. One idea is to allow a shared set of inducing inputs for all kernels (this could act as a form of regularasation)
Prediction with full covariance doesn't work because of numerical issuesOptimisation does not work- Gpflow bug: methods decorated with
autoflow
do not take named arguments