Releases: TuringLang/Bijectors.jl
v0.15.13
Bijectors v0.15.13
Exports extra functionality that should probably have been exported, namely ordered, isinvertible, and columnwise, from Bijectors.jl.
The docs have been thoroughly restructured.
Merged pull requests:
- Bump actions/checkout from 4 to 5 (#422) (@dependabot[bot])
- Enable Enzyme tests that are fixed now (#423) (@penelopeysm)
- enable Mooncake on 1.12 (#424) (@penelopeysm)
- CompatHelper: add new compat entry for Mooncake at version 0.4 for package test, (keep existing compat) (#425) (@github-actions[bot])
- overhaul docs and export some things that should be exported (#426) (@penelopeysm)
Closed issues:
- Dirichlet dimensionality off-by-one (#51)
- Improve Logit (#88)
- Model call:
no method matching eps(::Type{Real})when reusing aVarInfo(#109) - LKJ bijector is having problems (#144)
- Method error for Stacked (#169)
- Questions on custom bijectors (#234)
- Question on simplex bijector implementation (#283)
- Tests are failing for
VecCorrBijectorin very rare scenarios (#306) - Readme on the main page needs updated (#392)
orderedneeds an entry in the built docs + exporting, same withoutput_size(#396)- Update docs / workflow outputs to make running test instructions more clear (#397)
- PlanarLayer / Enzyme tests (#415)
v0.15.12
Bijectors v0.15.12
Improved implementation of the Enzyme rule for Bijectors.find_alpha.
Merged pull requests:
- Enzyme: migrate to easy_rule (#420) (@wsmoses)
- Disable the Enzyme rule test on 1.11 (#421) (@penelopeysm)
v0.15.11
Bijectors v0.15.11
Bijectors for ProductNamedTupleDistribution are now implemented.
Bijectors.output_size is now exported. This function provides information about the size of transformed variables. There are two main invocations:
output_size(b, input_size::Tuple)returns the size of the output ofb, given an input that has sizeinput_size.output_size(b, dist::Distribution)returns the size of the output ofb, given an input sampled from distributiondist. For most distributions this is implemented by callingoutput_size(b, size(dist)); however, ProductNamedTupleDistribution does not implementsize, so this method is necessary.
Merged pull requests:
- Refactor CI (#406) (@penelopeysm)
- Implement forward rule (frule!!) for find_alpha with integer arguments in Mooncake (#407) (@Copilot)
- Remove Tracker and Zygote references from tests and documentation (#409) (@Copilot)
- Add comprehensive GitHub Copilot instructions for Bijectors.jl development (#411) (@Copilot)
- Fix docs build by adding missing docstrings and reorganizing function documentation (#412) (@Copilot)
- Bump actions/checkout from 4 to 5 (#414) (@dependabot[bot])
- fix ci for 1.12 (#416) (@penelopeysm)
- Support ProductNamedTupleDistribution, export
output_size(#418) (@penelopeysm)
Closed issues:
- ReverseDiff doesn't support callable struct (#128)
- Rewrite: what do we want (#211)
- Zygote AD &
logpdffor transformed multivariate (#217) - Move all Enzyme related tests into a seprate CI (#347)
- Replace
ForwardDiff/ReverseDiff/Zygote/Mooncake/Enzymegradient call withDifferentiationInterface. (#373) - Remove
Tracker/Zygotefrom tests and docs. (#374) - Support ProductNamedTupleDistribution (#388)
- AD is troubled with Coupling layers (#401)
- Mooncake forward-mode (#404)
- ✨ Set up Copilot instructions (#410)
- PlanarLayer / Enzyme tests (#415)
v0.15.9
Bijectors v0.15.9
This patch fixes a precompilation error on BijectorsMooncakeExt for Mooncake > 0.4.147.
Merged pull requests:
- fixes for Mooncake 0.4.147 (#403) (@penelopeysm)
v0.15.8
Bijectors v0.15.8
This patch includes numerical stability fixes for LKJCholesky linking. It should make it much more possible to sample from LKJCholesky distributions with Hamiltonian Monte Carlo methods.
Merged pull requests:
- Cholesky numerical stability: Forward transform (#357) (@penelopeysm)
- CompatHelper: bump compat for AdvancedHMC to 0.8 for package test, (keep existing compat) (#389) (@github-actions[bot])
- Update README.md (#390) (@AoifeHughes)
- Remove Zygote dependency and update documentation for ForwardDiff integration (#393) (@AoifeHughes)
- Add buildkite pipeline file (#394) (@sunxd3)
v0.15.7
Bijectors v0.15.7
Added compatibility with ForwardDiff@1.
Merged pull requests:
- add ForwardDiff@1 (#378) (@penelopeysm)
- Replace mentions of stack with Stacked in docs (#381) (@mhauru)
- Enable doctests (#382) (@mhauru)
- CompatHelper: add new compat entry for Documenter at version 1 for package test, (keep existing compat) (#385) (@github-actions[bot])
- CompatHelper: add new compat entry for StableRNGs at version 1 for package test, (keep existing compat) (#386) (@github-actions[bot])
Closed issues:
- Error of Example in Documentation of Stacked (#380)
v0.15.6
Bijectors v0.15.6
Improved the numerical stability of the inverse transform for LKJCholesky
Merged pull requests:
- Implement logabsdetjac for Inverse{<:TruncatedBijector} for better numerical stability (#325) (@acertain)
- Cholesky numerical stability: inverse transform (#356) (@penelopeysm)
- CompatHelper: bump compat for AdvancedHMC to 0.7 for package test, (keep existing compat) (#369) (@github-actions[bot])
- Fix Enzyme interface (#372) (@yebai)
Closed issues:
- Document new interface (#34)
v0.15.5
Bijectors v0.15.5
This release added a new package extension to Bijectors.jl that is loaded when ReverseDiff and ChainRules are both loaded. This fixes a bug when using ReverDiff with Bijectors involving PDVecBijector.
Merged pull requests:
- Documentation and Turing Navigation CI improvement (#364) (@shravanngoswamii)
- Add
ChainRulestoBijectorsweak dependencies (#366) (@sunxd3) - Bump Julia compat to 1.10.8 (#368) (@penelopeysm)
Closed issues:
v0.15.4
Bijectors v0.15.4
Fix a bug that affected the Stacked bijector when using Zygote.
Merged pull requests:
Closed issues:
- Zygote is broken for Stacked transformations (#361)
v0.15.3
Bijectors v0.15.3
- Extended Zygote.jl compat entry to include 0.7
Merged pull requests:
- CompatHelper: bump compat for Zygote in [weakdeps] to 0.7, (keep existing compat) (#358) (@github-actions[bot])
- Use DOCUMENTER_KEY for CompatHelper (#360) (@penelopeysm)
Closed issues: