v0.5.66

Lux v0.5.66

Diff since v0.5.65

Merged pull requests:

• feat: NilArray for fast size propagation (#811) (@avik-pal)
• ci: setup benchmarking CI (#821) (@avik-pal)
• chore: bump minimum LuxTestUtils version (#850) (@avik-pal)
• docs: minor documentation changes (#855) (@avik-pal)
• chore: marking layers as deprecated (#856) (@avik-pal)
• chore(deps): bump crate-ci/typos from 1.23.6 to 1.24.1 (#857) (@dependabot[bot])

Closed issues:

• Continuous Benchmarking (#573)
• Lux.testmode not updating state for BatchNorm layers for nested models? (#849)
• Add Float128 support (#851)
• Add multiple cpu cores and multiple Julia computers support (#852)
• Enzyme.Forward hits Octavian dispatch in Dense (#853)

v0.5.65

Lux v0.5.65

Diff since v0.5.64

Merged pull requests:

• feat: improve batched jacobian (#848) (@avik-pal)

v0.5.64

Lux v0.5.64

Diff since v0.5.63

Merged pull requests:

• test: remove Flux as a direct test dep (#838) (@avik-pal)
• test: remove some of the unnecessary Flux tests (#839) (@avik-pal)
• refactor: cleanup of internals (#840) (@avik-pal)
• fix: remove type pirated functions from Lux (#843) (@avik-pal)
• chore(deps): bump actions/upload-artifact from 2 to 4 (#844) (@dependabot[bot])
• chore(deps): bump crate-ci/typos from 1.23.5 to 1.23.6 (#845) (@dependabot[bot])
• CompatHelper: add new compat entry for Static at version 1 for package test, (keep existing compat) (#846) (@github-actions[bot])

Closed issues:

• how to use Lux.jl utility 'BinaryCrossEntropy' (#841)
• Mixed-Precision Matrix Multiply Performance Regression (#847)

v0.5.63

Lux v0.5.63

Diff since v0.5.62

Merged pull requests:

• docs: add new function to docs (#813) (@avik-pal)
• fix: update Dynamic Expressions to 0.19 (#814) (@avik-pal)
• docs: add documentation for MLDataDevices (#815) (@avik-pal)
• CompatHelper: add new compat entry for MLDataDevices at version 1 for package docs, (keep existing compat) (#818) (@github-actions[bot])
• test: try separating the test Project files (#819) (@avik-pal)
• feat: use faster version of batched matmul (#820) (@avik-pal)
• ci: add CI to benchmark load times (#822) (@avik-pal)
• chore(deps): bump actions/checkout from 2 to 4 (#823) (@dependabot[bot])
• chore(deps): bump peter-evans/create-or-update-comment from 3 to 4 (#824) (@dependabot[bot])
• chore(deps): bump julia-actions/setup-julia from 1 to 2 (#825) (@dependabot[bot])
• chore(deps): bump peter-evans/find-comment from 2 to 3 (#826) (@dependabot[bot])
• chore(deps): bump julia-actions/cache from 1 to 2 (#827) (@dependabot[bot])
• fix: mark objective function as Const (#835) (@avik-pal)
• ci: separate testing for groups in buildkite (#836) (@avik-pal)
• chore: update all AMDGPU compats (#837) (@avik-pal)

Closed issues:

• NCCL Complex wrapper (#747)
• Gradient checkpointing/ rematerialization (#816)

v0.5.62

Lux v0.5.62

Diff since v0.5.61

Merged pull requests:

• Allow 2d input if RNN order is BatchLastIndex (#778) (@NeroBlackstone)
• test: remove @test_nowarn testing (#781) (@avik-pal)
• fix: don't reuse pullback for safety (#782) (@avik-pal)
• improvements to compact macro (#783) (@avik-pal)
• test: warp @inferred with @test (#784) (@avik-pal)
• chore: add NNlib as a direct dep (#785) (@avik-pal)
• fix: update to latest LuxLib API + deprecations (#786) (@avik-pal)
• perf: fix enzyme benchmarks (#787) (@avik-pal)
• test: trigger enzyme tests (#788) (@avik-pal)
• docs: fix typo in "JVP & VJP Wrappers" (#789) (@ldeso)
• docs: update docs from downstream changes (#790) (@avik-pal)
• CompatHelper: bump compat for WeightInitializers to 1, (keep existing compat) (#791) (@github-actions[bot])
• CompatHelper: bump compat for WeightInitializers to 1 for package docs, (keep existing compat) (#792) (@github-actions[bot])
• test: improved testing (#793) (@avik-pal)
• feat: improvements to the Training API (#794) (@avik-pal)
• feat: easy mechanism to set preferences (#798) (@avik-pal)
• CompatHelper: bump compat for SymbolicUtils to 3 for package SymbolicOptimalControl, (keep existing compat) (#799) (@github-actions[bot])
• test: update to the newer LuxTestUtils (#800) (@avik-pal)
• chore(deps): bump crate-ci/typos from 1.23.2 to 1.23.5 (#804) (@dependabot[bot])
• refactor: move TrackerExt in a directory (#806) (@avik-pal)

Closed issues:

• Allow "const" arrays as inputs to @compact (#588)
• Meta-Issue for improvements to @compact (#606)
• Add type-stability checks via DispatchDoctor.jl (#683)
• Feature Request: Allow recurrent layers with 2D input (features * seq_length), even if the order is BatchLastIndex (#767)

v0.5.61

Lux v0.5.61

Diff since v0.5.60

Merged pull requests:

• refactor: bring Optimisers.jl into main deps (#754) (@avik-pal)
• refactor: drop the AMDGPU extension (#755) (@avik-pal)
• rearrange code in extensions (#756) (@avik-pal)
• fix: use proper qualified accesses for modules (#757) (@avik-pal)
• docs: remove redundant old preferences (#759) (@avik-pal)
• feat: allow multiple @return (#760) (@avik-pal)
• Making all eltypes Float32 in Fitting a Polynomial using MLP (#761) (@Sleort)
• docs: fix inline math rendering (#762) (@avik-pal)
• refactor: use the faster get_device_type (#763) (@avik-pal)
• refactor: move ForwardDiff.jl into main deps (#764) (@avik-pal)
• test: set st to training (#765) (@avik-pal)
• chore(deps): bump crate-ci/typos from 1.23.1 to 1.23.2 (#766) (@dependabot[bot])
• Update docstring dropout (#770) (@dmetivie)
• chore: recommend GH Discussions for Q/A (#774) (@avik-pal)

Closed issues:

• Broadcast Layer (#426)
• Add improper qualified accesses checks (#667)
• Feature request: TimeDistributed Layer (#758)

v0.5.60

Lux v0.5.60

Diff since v0.5.59

Merged pull requests:

• Bidirectional RNN (#708) (@NeroBlackstone)
• Bump crate-ci/typos from 1.22.9 to 1.23.1 (#748) (@dependabot[bot])
• chore: bump simplechains version (#749) (@avik-pal)
• CompatHelper: bump compat for SciMLSensitivity to 7 for package NeuralODE, (keep existing compat) (#750) (@github-actions[bot])
• docs: restructure the manual entries a bit (#751) (@avik-pal)

Closed issues:

• Feature request: Bidirectional for RNN layer. (#687)

v0.5.59

Lux v0.5.59

Diff since v0.5.58

Merged pull requests:

• fix: nested ad when using direct eval in function (#745) (@avik-pal)
• CompatHelper: add new compat entry for GPUArraysCore at version 0.1 for package docs, (keep existing compat) (#746) (@github-actions[bot])

Closed issues:

• Error in compute_gradients when loss already has a Zygote.gradient (#743)

v0.5.58

Lux v0.5.58

Diff since v0.5.57

Merged pull requests:

• CompatHelper: bump compat for AMDGPU in [weakdeps] to 0.9, (keep existing compat) (#717) (@github-actions[bot])
• Makes name concrete types (#723) (@avik-pal)
• CompatHelper: add new compat entry for StaticArrays at version 1 for package docs, (keep existing compat) (#724) (@github-actions[bot])
• CompatHelper: add new compat entry for KernelAbstractions at version 0.9 for package docs, (keep existing compat) (#725) (@github-actions[bot])
• Bump crate-ci/typos from 1.22.7 to 1.22.9 (#726) (@dependabot[bot])
• Performance Pitfalls and How to Catch them (#727) (@avik-pal)
• CompatHelper: bump compat for DynamicExpressions in [weakdeps] to 0.18, (keep existing compat) (#728) (@github-actions[bot])
• CompatHelper: bump compat for DynamicExpressions to 0.18 for package SymbolicOptimalControl, (keep existing compat) (#729) (@github-actions[bot])
• Store the optimizer in TrainState (#731) (@avik-pal)
• Simply show implementations and make them round-trippable (#732) (@avik-pal)
• Try removing the type assert with this (#734) (@avik-pal)
• Add enzyme support for loss functions from LossFunctions.jl (#736) (@avik-pal)
• Mark cartersian index tests on cuda broken for now (#737) (@avik-pal)
• Run CI on pre (#739) (@avik-pal)
• Revert bee2de7-1188db7 (#740) (@avik-pal)
• Use shorthand syntax of @concrete (#741) (@avik-pal)
• Check status of broken tests (#742) (@avik-pal)

Closed issues:

• Auto detect and warn against performance pitfalls (#699)
• Inference using NN :: Chain inside a GPU kernel (#720)
• custom show is often not valid julia syntax to reconstruct (#730)

v0.5.57

Lux v0.5.57

Diff since v0.5.56

Merged pull requests:

• Try wider docs (#721) (@avik-pal)
• Compiled ReverseDiff for training on CPU (#722) (@avik-pal)

Closed issues:

• Add a compiled tape version for ReverseDiff (#642)