DifferentiationInterface-v0.5.10
·
120 commits
to main
since this release
TLDR
- Improvements in Hessian performance with
SecondOrder(AutoForwardDiff(), AutoReverseDiff()) - Some behind-the-scenes work to enable differentiation of more generic structs, with the first tests for Flux and Lux neural nets (not part of API guarantees)
What's Changed
- CompatHelper: add new compat entry for FiniteDifferences in [weakdeps] at version 0.12 for package DifferentiationInterfaceTest, (keep existing compat) by @github-actions in #368
- CompatHelper: add new compat entry for Flux in [weakdeps] at version 0.14 for package DifferentiationInterfaceTest, (keep existing compat) by @github-actions in #367
- Fix Flux compat bound by @gdalle in #369
- Smarter symmetric decompression by @gdalle in #370
- Debug Flux tests by @gdalle in #371
- Fix size-1 Jacobian with Enzyme by @gdalle in #374
- Handle closures with Enzyme by @gdalle in #375
- Relax DocStringExtensions compat entries by @adrhill in #379
- Keep tests in same env by @gdalle in #381
- Better handling of
AutoEnzyme(constant_function=false)by @gdalle in #382 - Put some Enzyme stuff in utils by @gdalle in #384
- Test ForwardDiff over ReverseDiff by @gdalle in #386
- ForwardDiff over anything - prepare inner gradient for HVP by @gdalle in #385
- Update Compat's compat by @gdalle in #389
- Bump SparseConnectivityTracer compat to
v0.6by @adrhill in #392 - Don't test on 'lts' and 'pre' for draft PRs by @gdalle in #391
- Add Lux tests by @gdalle in #372
- Revert handling of
AutoEnzyme(; constant_function)by @gdalle in #394 - Bump DI version to 0.5.10 by @gdalle in #395
Full Changelog: DifferentiationInterface-v0.5.9...DifferentiationInterface-v0.5.10
Contributors
adrhill and gdalle