AugmentedGaussianProcesses.jl

AugmentedGaussianProcesses v0.8.2

Diff since v0.8.1

Merged pull requests:

• CompatHelper: bump compat for "MCMCChains" to "4.0" (#49) (@github-actions[bot])
• CompatHelper: bump compat for "Zygote" to "0.5" (#50) (@github-actions[bot])
• Fix different issues with GP regression (#51) (@theogf)

AugmentedGaussianProcesses v0.8.1

Diff since v0.8.0

Closed issues:

• Problem with variational student t processes (#42)

Merged pull requests:

• Refactoring and completeness of the tests (#40) (@theogf)
• Fixed issues with VStP (#43) (@theogf)
• CompatHelper: bump compat for "Clustering" to "0.14" (#44) (@github-actions[bot])
• CompatHelper: bump compat for "KernelFunctions" to "0.4" (#45) (@github-actions[bot])
• CompatHelper: bump compat for "PDMats" to "0.10" (#46) (@github-actions[bot])
• CompatHelper: bump compat for "Distances" to "0.9" (#47) (@github-actions[bot])

AugmentedGaussianProcesses v0.8.0

Diff since v0.7.1

Closed issues:

• StackOverflow in Examples Script (#20)

Merged pull requests:

• CompatHelper: bump compat for "StatsBase" to "0.33" (#38) (@github-actions[bot])

AugmentedGaussianProcesses v0.7.1

Diff since v0.7.0

Closed issues:

• Fix the inverse laplace transform sampler algorithm (#36)

Merged pull requests:

• CompatHelper: bump compat for "Distributions" to "0.23" (#37) (@github-actions[bot])

AugmentedGaussianProcesses v0.7.0

Diff since v0.6.0

Closed issues:

• Work on kernel module (#8)
• Improve coverage (#15)
• Duplicate compat entry (#33)

Merged pull requests:

• Install TagBot as a GitHub Action (#22) (@JuliaTagBot)
• CompatHelper: bump compat for "MCMCChains" to "2.0" (#23) (@github-actions[bot])
• CompatHelper: add new compat entry for "Distances" at version "0.8" (#24) (@github-actions[bot])
• CompatHelper: bump compat for "StatsFuns" to "0.9" (#25) (@github-actions[bot])
• CompatHelper: bump compat for "Distributions" to "0.22" (#26) (@github-actions[bot])
• CompatHelper: add new compat entry for "Zygote" at version "0.4" (#27) (@github-actions[bot])
• CompatHelper: add new compat entry for "SimpleTraits" at version "0.9" (#28) (@github-actions[bot])
• CompatHelper: add new compat entry for "Reexport" at version "0.2" (#29) (@github-actions[bot])
• CompatHelper: add new compat entry for "KernelFunctions" at version "0.2" (#30) (@github-actions[bot])
• CompatHelper: add new compat entry for "DataStructures" at version "0.17" (#31) (@github-actions[bot])
• CompatHelper: bump compat for "MCMCChains" to "3.0" (#32) (@github-actions[bot])

Large refactoring of the package:

• Variational and kernel parameters are contained into _VGP and _SVGP.
• The package now relies on KernelFunctions.jl allowing to use automatic differentation, also the derivatives are not adapted for it yet (WIP).
• AffineMean has been added to have a parametric mean prior (trainable as well)
• MOSVGP a multi-output model has been added to make a linear combination of inducing points see work of Pablo Moreno-Munoz
• More likelihoods have been added (Matern, NegativeBinomial)
• MCGP a sampling based Gaussian Process, for now only Gibbs Sampling is available but it is planned to use HMC via AdvancedHMC.jl
• Custom likelihoods will be available by simply giving the pdf function (WIP)
• NumericalVI is temporarily not working this will be fixed soon

v0.6.0 (2019-11-19)

Diff since v0.5.1

Merged pull requests:

• Added Negative Binomial Likelihood and Fixed Poisson (#19) (theogf)

Bug fixes and preparations for new general methods

• Improved predictions using quadrature
• Draft for Student-T Processes (beta)
• Gibbs Sampling for Laplace likelihood (and Bayesian SVM soon)
• Adapted tests

v0.5.1 (2019-09-02)

Diff since v0.5.0

v0.5.0 (2019-08-01)

• Added variational inference via quadrature
• Corrected a lot of bugs
• Unified the code structure

Diff since v0.4.3

Closed issues:

• Add all possible likelihoods (#6)

This release includes

• Nicer training output via ProgressMeter.jl
• More easily modifiable optimizer for the hyperparameters and inducing point locations : keywords optimizer and Zoptimizer
• Heteroscedastic likelihood in beta version
• More detailed documentation

v0.4.3 (2019-07-01)

Diff since v0.4.2

Merged pull requests:

• Correct comparison.md (#17) (thomaspinder)

Corrected Laplace ELBO
Added Poisson Likelihood for events datasets
Custom mean priors : ZeroMean, ConstantMean, EmpiricalMean (see docs)

Simplification of the likelihood and inference struct names
Documentation improvement

Complete remaking of the package

• Now only 3 models : GP, VGP and SVGP parametrized on their likelihood and inference
• Unique background for all likelihoods and inference to simplify the addition of new likelihoods
• Multioutput GPs (still some bugs there), but does not allow dependency between outputs yet
• Functions are more julian : train!(model), predict_f(model,X_test)
• Automatic plotting of the models (for 1D and 2D inputs) via Makie.plot(model)
• Sampling from the model via rand
• Multiclass likelihood is now fully functionning with multiple features described in the docs

Nothing new, just removed the integrated GradDescent

More complete documentation with full examples.
More stable predictions by replacing QuadGK.jl with Expectations.jl (but slightly slower)
Documentation improved for many functions

No major improvements

New major update,
With the new Student-T Likelihood added
Possibility to save and load trained models
Important optimization for the matrix and hyperoptimization side
GradDescent integrated to the package (temporarily)

Added a lot of documentation
Better testing
Reconstruction of the structure
First drafts of Multiclass