v0.2.12

AdvancedVI v0.2.12

Diff since v0.2.11

This release has been identified as a backport.
Automated changelogs for backports tend to be wildly incorrect.
Therefore, the list of issues and pull requests is hidden.

v0.4.0

AdvancedVI v0.4.0

Diff since v0.3.2

changelog

• Added proximal operator for the location-scale family proposed by ¹. This proximal operator is supported for the optimization rules DoG, DoWG, Descent.

Merged pull requests:

• Proximal operator for the entropy of location-scale families (#168) (@Red-Portal)

1. Domke, Justin. "Provable smoothness guarantees for black-box variational inference." International Conference on Machine Learning. PMLR, 2020. ↩

v0.3.2

AdvancedVI v0.3.2

Diff since v0.3.1

Merged pull requests:

• try rm ChainRulesCore (#167) (@yebai)
• CompatHelper: bump compat for ForwardDiff to 1 for package docs, (keep existing compat) (#171) (@github-actions[bot])
• bump ForwardDiff v1.0 (#172) (@Red-Portal)

Closed issues:

• Missing API method (#32)

v0.3.1

AdvancedVI v0.3.1

Diff since v0.3.0

Major Changes:

• Previously, AdvancedVI directly called Enzyme through its official public interface. Now, t is called through DifferentiationInterface.
• All AD calls are made with prepare_gradient.

Merged pull requests:

• CompatHelper: add new compat entry for AdvancedVI at version 0.3 for package bench, (keep existing compat) (#154) (@github-actions[bot])
• CompatHelper: bump compat for Zygote to 0.7 for package bench, (keep existing compat) (#156) (@github-actions[bot])
• CompatHelper: bump compat for Zygote to 0.7 for package test, (keep existing compat) (#157) (@github-actions[bot])
• Documentation and Turing Navigation CI improvement (#158) (@shravanngoswamii)
• remove unnecessary weakdeps (#159) (@yebai)
• simplify package extension for post Julia 1.10 (#161) (@yebai)
• Update Enzyme usage in test files (#162) (@yebai)
• Optimize code for better performance and maintainability (#163) (@yebai)
• Fix ClipScale dispatch error (#164) (@yebai)
• Remove the Enzyme extension, prepare gradient (#166) (@yebai)

Closed issues:

• Enable prepare_gradient for DifferentiationInterface (#101)
• Do we still need the Enzyme extension with DI? (#160)

v0.3.0

AdvancedVI v0.3.0

Diff since v0.2.11

Breaking changes

• Complete rewrite of AdvancedVI with major changes in the API. (Refer to general usage and the example.)

New Features

• Added full-rank and low-rank covariance low-rank variational families. (See the docs.)
• Added the sticking-the-landing control variate. (See the docs.)
• Added the score gradient estimator of the ELBO gradient with the leave-one-out control variate (also known as VarGrad)
• Added parameter averaging. (See the docs)
• Added parameter-free optimization algorithms. (See the docs)

Merged pull requests:

• Minor Touches for ScoreGradELBO (#99) (@Red-Portal)
• Drop support for pre-1.10 (#129) (@Red-Portal)
• refactor interface for projections/proximal operators (#147) (@Red-Portal)
• CompatHelper: bump compat for AdvancedVI to 0.2 for package docs, (keep existing compat) (#153) (@github-actions[bot])

Closed issues:

• Set up unit tests for GPU support (#70)

v0.2.11

AdvancedVI v0.2.11

Diff since v0.2.10

Merged pull requests:

• CompatHelper: bump compat for Bijectors in [weakdeps] to 0.14, (keep existing compat) (#137) (@github-actions[bot])
• CompatHelper: bump compat for Bijectors to 0.14 for package bench, (keep existing compat) (#138) (@github-actions[bot])
• CompatHelper: bump compat for Bijectors to 0.14 for package test, (keep existing compat) (#139) (@github-actions[bot])
• CompatHelper: bump compat for Bijectors to 0.14 for package docs, (keep existing compat) (#140) (@github-actions[bot])
• bump bijectors on v0.2 (#152) (@Red-Portal)

v0.2.10

AdvancedVI v0.2.10

Diff since v0.2.9

Merged pull requests:

• Add compatibility for Bijectors 0.14 (#136) (@penelopeysm)

v0.2.9

AdvancedVI v0.2.9

Diff since v0.2.8

Merged pull requests:

• Add benchmarks (#104) (@Red-Portal)
• Cleanup empty extensions (#128) (@Red-Portal)
• CompatHelper: bump compat for Documenter to 1 for package docs, (keep existing compat) (#130) (@github-actions[bot])
• bump Enzyme version in v0.2 (#132) (@Red-Portal)
• Bump patch version to 0.2.9 (#135) (@mhauru)

Closed issues:

• are there any special purposes of having empty pkg extensions? (#100)

v0.2.8

AdvancedVI v0.2.8

Diff since v0.2.7

Merged pull requests:

• Copy docs workflows to v0.2 branch (#107) (@penelopeysm)

v0.2.7

AdvancedVI v0.2.7

Diff since v0.2.6

Merged pull requests:

• Rewriting advancedvi docs (#53) (@Red-Portal)
• fix rename adbackend to adtype (#60) (@Red-Portal)
• add continuous benchmarking (#61) (@Red-Portal)
• remove signature with user-defined restructure (#64) (@Red-Portal)
• add indirection for update step, add projection for LocationScale (#65) (@Red-Portal)
• fix avoid re-defining the differentiation objective to support AD pre-compilation (#66) (@Red-Portal)
• Enable Enzyme (#67) (@Red-Portal)
• Add ScoreELBO objective (#72) (@arnauqb)
• Create Format.yml (#73) (@shravanngoswamii)
• Add "low-rank" variational families (#76) (@Red-Portal)
• Run benchmarks only on PR (#77) (@shravanngoswamii)
• Exclude benchmarks from DocsNav (#78) (@shravanngoswamii)
• Fix Enzyme extension (#79) (@devmotion)
• Formatting all files (#80) (@yebai)
• Parameter-Free Optimization Algorithms (#81) (@Red-Portal)
• Tapir.jl Integration (#86) (@willtebbutt)
• CompatHelper: add new compat entry for Tapir in [weakdeps] at version 0.2, (keep existing compat) (#87) (@github-actions[bot])
• Fix Enzyme tests on older Julia versions (#88) (@Red-Portal)
• Grant push permissions to Benchmark.yml (#90) (@penelopeysm)
• Only upload benchmark results if not on a fork (#91) (@penelopeysm)
• Adapt to pending Enzyme breaking change (#92) (@wsmoses)
• Revert "Adapt to pending Enzyme breaking change" (#93) (@Red-Portal)
• Revert "Revert "Adapt to pending Enzyme breaking change"" (#94) (@Red-Portal)
• Add SSH key for all relevant workflows (#97) (@penelopeysm)
• Migrate to DifferentiationInterface (#98) (@Red-Portal)
• add subdirectories to compathelper (#102) (@Red-Portal)
• Add minimal documentation for v0.2 (#105) (@penelopeysm)
• Create a new release so that stable docs get generated (#106) (@penelopeysm)

Closed issues:

• Stein Variational Gradient Descent (SVGD) (#3)
• Double Stochasticity (#4)
• Callback function during training (#5)
• Rethinking AdvancedVI (#24)
• Setting up Documenter (#39)
• Add Tapir to AD tests (#63)
• Set up JuliaFormatter (#69)
• Run benchmarking only on PRs (#74)
• var and cov on MvLocationScale secretly assume the base distribution is standardized (#75)
• Where are the benchmarks posted? (#82)
• Make use of DifferentiationInterface.jl? (#85)