v0.6.0

AdvancedVI v0.6.0

Diff since v0.5.0

New Algorithms

This update adds new variational inference algorithms in light of the flexibility added in the v0.5 update.
Specifically, the following measure-space optimization algorithms have been added:

• KLMinWassFwdBwd
• KLMinNaturalGradDescent
• KLMinSqrtNaturalGradDescent
• FisherMinBatchMatch

Interface Change (breaking)

The objective value returned by estimate_objective is now the value to be minimized by the algorithm.
For instance, for ELBO maximization algorithms, estimate_objective will return the negative ELBO.

Behavior Change

In addition, KLMinRepGradDescent, KLMinRepGradProxDescent, KLMinScoreGradDescent will now throw a RuntimeException if the objective value estimated at each step turns out to be degenerate (Inf or NaN). Previously, the algorithms ran until max_iter even if the optimization run had failed.

Merged pull requests:

• Move general reshuffling stuff into a separate file (#208) (@Red-Portal)
• Move SubsampledNormals test target to its own file under test/models/ (#209) (@Red-Portal)
• Add the forward-backward Wasserstein Gaussian variational inference algorithm (#210) (@Red-Portal)
• Add natural gradient variational inference algorithms (#211) (@Red-Portal)
• Check that the objective value is finite in the shared step (#212) (@Red-Portal)
• Refactor algorithm unit tests (#213) (@Red-Portal)
• Add documentation for the natural gradient algorithms (#214) (@Red-Portal)
• Specify convention for estimate_objective (#215) (@Red-Portal)
• CompatHelper: bump compat for AdvancedVI to 0.5 for package bench, (keep existing compat) (#216) (@github-actions[bot])
• CompatHelper: bump compat for AdvancedVI to 0.5 for package docs, (keep existing compat) (#217) (@github-actions[bot])
• Batch-and-Match algorithm for minimizing the covariance-weighted Fisher divergence (#218) (@Red-Portal)

Closed issues:

• Natural Gradients + Monte Carlo VI (#1)