v0.3.0

Added implementation with Keras functionality.

1. Training fit model.fit()
2. Initialising and saving models with Keras serialisation

v0.2.0

1. Changed implementation of SVGP Gating Network:
   • SVGPGatingFunction, SVGPGatingNetworkBinary and SVGPGatingNetworkMulti are now depreciated as their functionality has been moved to a single class (SVGPGatingNetwork) with different likelihoods,
     • SVGPGatingNetwork now inherits from SVGPModel that accepts either a Bernoulli or Softmax likelihood, governing the number of gating functions/experts to use.
2. Implemented different lower bounds for comparison and now defaults to a better lower bound,
   • This bound corresponds to a likelihood approximation that breaks the marginalisation over exponentially many sets of assignments.
3. The examples directory trains and compares the bounds on the mcycle and quadcotper data sets.
4. Updated mcycle notebooks to work with new gating network and lower bounds.
5. Removed maths from docs as it's outdated and incorrect. The plan is to add a link to thesis or paper once published.

Release 0.1.0

updated tf/tfp/gpflow

Release 0.1.0

Initial release with a working implementation of Mixtures of Sparse Variational Gaussian Process Experts.

Four examples on different data sets are shown in the examples/ directory.

Code is documented on Read The Docs.