v1.5.0

Lux v1.5.0

Diff since v1.4.4

Merged pull requests:

• CompatHelper: add new compat entry for ProgressTables at version 0.1 for package CIFAR10, (keep existing compat) (#1154) (@github-actions[bot])
• device(NN) should only give warning once (#1156) (@vpuri3)
• feat: conditional VAE (#1157) (@avik-pal)
• fix: ConditionalVAE on CI (#1159) (@avik-pal)
• CompatHelper: add new compat entry for Optimisers at version 0.4 for package ConditionalVAE, (keep existing compat) (#1161) (@github-actions[bot])
• CompatHelper: add new compat entry for StableRNGs at version 1 for package ConditionalVAE, (keep existing compat) (#1162) (@github-actions[bot])
• CompatHelper: add new compat entry for Comonicon at version 1 for package ConditionalVAE, (keep existing compat) (#1163) (@github-actions[bot])
• [MLDataDevices] Bump deps (#1164) (@pxl-th)
• docs: migrate most examples to Reactant (#1180) (@avik-pal)
• chore: bump crate-ci/typos from 1.28.4 to 1.29.4 (#1183) (@dependabot[bot])
• fix: pass in RNG to shuffle (#1188) (@avik-pal)
• feat: allow no grad option for reactant (#1190) (@avik-pal)

Closed issues:

• Recurrent cells cannot be chained with other layers (#1155)
• No Grad option for TrainState single_train_step(!) (#1181)
• sparse_init doesn't use provided rng fully (#1185)
• Incorrect IR generated for some neural networks (#1186)
• WeightInitializers.DeviceAgnostic doesn't respect Reactant (#1187)
• Export utilities in partial.jl (#1189)

WeightInitializers-v1.1.0

WeightInitializers WeightInitializers-v1.1.0

Diff since WeightInitializers-v1.0.5

Merged pull requests:

• docs: migrate most examples to Reactant (#1180) (@avik-pal)

Closed issues:

• WeightInitializers.DeviceAgnostic doesn't respect Reactant (#1187)

WeightInitializers-v1.0.5

WeightInitializers WeightInitializers-v1.0.5

Merged pull requests:

• Rewrite (#7) (@avik-pal)
• Rename to Lux (#11) (@avik-pal)
• Initial Documentation (#14) (@avik-pal)
• Minor Updates (#15) (@avik-pal)
• Better CUDNN Dispatches (#16) (@avik-pal)
• Tutorials (#21) (@avik-pal)
• Proper dispatch for types not supported by CUDNN (#23) (@avik-pal)
• [WIP] Recurrent Neural Networks (#24) (@avik-pal)
• Fix math display in docs (#27) (@gdalle)
• Initial ViT Implementation & Pretrained ImageNet Models (#29) (@avik-pal)
• CompatHelper: bump compat for Setfield to 1, (keep existing compat) (#30) (@github-actions[bot])
• Code Formatting -- SciMLStyle (#31) (@avik-pal)
• Cleanup generated function style (#33) (@avik-pal)
• Update README.md (#37) (@zsz00)
• Fix doc for PairwiseFusion (#39) (@theabhirath)
• Extending Scale to allow for multiple dimension inputs (#40) (@theabhirath)
• Fix Zygote error caused due to fill! (#41) (@theabhirath)
• CompatHelper: bump compat for ComponentArrays to 0.12, (keep existing compat) (#43) (@github-actions[bot])
• Update JET tests to allow julia v1.6 (#47) (@avik-pal)
• Formatting updates and relax parameter type (#48) (@avik-pal)
• Enable doctests in CI (#51) (@avik-pal)
• fix quickstart example (#52) (@visr)
• Test on 1.8 (#54) (@avik-pal)
• Separate out testing unreleased julia versions (#55) (@avik-pal)
• Cleaner and Better Documentation (#56) (@avik-pal)
• Bump Pkg Compats (#66) (@avik-pal)
• CompatHelper: bump compat for MLDatasets to 0.7 for package examples, (keep existing compat) (#67) (@github-actions[bot])
• Manual to translate Flux to Lux (#69) (@avik-pal)
• Try codecov for doctests (#70) (@avik-pal)
• Add tests for utility functions (#74) (@avik-pal)
• Add tip to install packages (#76) (@Karthik-d-k)
• More Testing + Deprecate Nonsensical Functions + Better Naming for Kwargs (#80) (@avik-pal)
• CompatHelper: add new compat entry for Optimisers at version 0.2, (keep existing compat) (#82) (@github-actions[bot])
• Update rrules so that we can support Yota (#85) (@avik-pal)
• CompatHelper: bump compat for FluxMPI to 0.6 for package examples, (keep existing compat) (#86) (@github-actions[bot])
• Update comparison section in overview.md (#88) (@ToucheSir)
• Fix typos (#89) (@claforte)
• Fix minor typos in the docs (#93) (@gabrevaya)
• making x Float32 in migrate from Flux example (#97) (@gabrevaya)
• add init_hidden_state function (#101) (@gabrevaya)
• JLArray is now registered (#103) (@YichengDWu)
• [LuxTraining] Wrappers for less clunky training loops (#104) (@avik-pal)
• Use OneHotArrays (#105) (@YichengDWu)
• Fixes WeightNorm with zero Parameter bug (#106) (@avik-pal)
• fix state update in NeuralODE example (#107) (@gabrevaya)
• Deprecate elementwise_* and applyactivation (#113) (@avik-pal)
• Go through the dense bias deprecation (#114) (@avik-pal)
• Fix Scale's paramlength (#116) (@lungd)
• Trainable hidden states (#117) (@lungd)
• Rnn bias deprecation (#120) (@lungd)
• Add use_bias kwarg to LSTMCell and GRUCell (#121) (@lungd)
• Update docs for dense layer (#124) (@avik-pal)
• Upper bound ComponentArrays (#125) (@avik-pal)
• Relax ComponentArrays compat (#126) (@avik-pal)
• Layer Normalization Implementation (#127) (@avik-pal)
• LSTM docs: don't go over first element in sequence twice (#132) (@visr)
• fix PairwiseFusion docs (#133) (@YichengDWu)
• Generic recurrent cells (#136) (@jumerckx)
• relu tests with finite diff is too unreliable (#137) (@avik-pal)
• Add kaiming initialization (#138) (@YichengDWu)
• Remove Val in typeinfo of WeightNorm (#140) (@avik-pal)
• Named Layers inside Generic Containers (#143) (@avik-pal)
• Allow fmapping over the model (#144) (@avik-pal)
• Update Imagenet example (#147) (@avik-pal)
• Make normalization more AD friendly (Diffractor) (#148) (@avik-pal)
• Fix CuArray -> Array rrule (#149) (@avik-pal)
• Allow indexing into Chains (#150) (@avik-pal)
• API for freezing layers (#151) (@avik-pal)
• Allow controlling fast activation transformation (#153) (@avik-pal)
• Introducing LuxLib.jl: Effectively pullout some of the custom layer implementations from Lux.jl (#154) (@avik-pal)
• Try relaxing JET version (#155) (@avik-pal)
• Update to use LuxLib (#156) (@avik-pal)
• Allow dispatch using Lux.apply (#158) (@avik-pal)
• Mark non differentiable code paths (#160) (@avik-pal)
• Fix generic GN dispatch for non 4D arrays (#161) (@avik-pal)
• Add dispatch for subarray (#162) (@avik-pal)
• Add More Layers (#163) (@avik-pal)
• Fix type stability in normalization implementation (#164) (@avik-pal)
• Codecov for lib directories Take 2 (#165) (@avik-pal)
• Add freeze tests to runtests (#166) (@avik-pal)
• Precompile common workflows + check invalidations (#167) (@avik-pal)
• Make normalization typestable (#168) (@avik-pal)
• Add a manual page on precompilation (#169) (@avik-pal)
• Deprecate Lux.transform in favor of Flux2Lux.jl (#170) (@avik-pal)
• Remove dead code and improve var for Tracker.jl support (#171) (@avik-pal)
• Hyper Network Example (#172) (@avik-pal)
• Modify mkdocs settings (#173) (@avik-pal)
• Make ViT work on GPUs (#174) (@avik-pal)
• Add sensible recurrent layer wrappers (#175) (@avik-pal)
• setup only on AbstractRules (#176) (@avik-pal)
• Start using Flux2Lux (#177) (@avik-pal)
• Fix some displays (#178) (@avik-pal)
• Relax dropout types (#179) (@avik-pal)
• Add instancenorm and alpha_dropout implementations (#180) (@avik-pal)
• Add InstanceNorm and AlphaDropout (#181) (@avik-pal)
• CompatHelper: bump compat for MLUtils to 0.3 for package examples, (keep existing compat) (#184) (@github-actions[bot])
• remove convert rrule (#185) (@ArnoStrouwen)
• CompatHelper: bump compat for OneHotArrays to 0.2 for package examples, (keep existing compat) (#186) (@github-actions[bot])
• CompatHelper: bump compat for Turing to 0.22 for package examples, (keep existing compat) (#188) (@github-actions[bot])
• Fix layer_map for custom layers (#189) (@avik-pal)
• add example of DDIM implementation (#190) (@yng87)
• LuxCore.jl: Extremely light dependency for Lux Compatibility (#191) (@avik-pal)
• Revert github workflows for merged LuxCore.jl (#193) (@avik-pal)
• CompatHelper: bump compat for MLUtils to 0.3 for package ImageNet, (keep existing compat) (#194) (@github-actions[bot])
• CompatHelper: bump compat for Setfield to 1 for package ImageNet, (keep existing compat) (#195) (@github-actions[bot])
• CompatHelper: bump compat for OneHotArrays to 0.2 for package ImageNet, (keep existing compat) (#196) (@github-actions[bot])
• ADAM -> Adam (#197) (@cossio)
• CompatHelper: bump compat for Functors to 0.4, (keep existing compat) (#199) (@github-actions[bot])
• CompatHelper: bump compat for Functors to 0.4 for package examples, (keep existing compat) (#200) (@github-actions[bot])
• CompatHelper: bump compat for Functors to 0.4 for package ImageNet, (keep existing compat) (#201) (@github-actions[bot])
• Add easy tied weights/parameter sharing support (#202) (@avik-pal)
• CompatHelper: bump compat for Functors to 0.4 for package LuxCore, (keep existing compat) (#203) (@github-actions[bot])
• CompatHelper: add new compat entry for Zygote at version 0.6 for package DDIM, (keep existing compat) (#218) (@github-actions[bot])
• Update DDIM compat requirements (#219) (@avik-pal)
• Update examples (#221) (@avik-pal)
• CompatHelper: bump compat for Turing to 0.23 for package examples, (keep existing compat) (#222) (@github-actions[bot])
• Fix docs (#223) (@avik-pal)
• CompatHelper: bump compat for MLUtils to 0.4 for package examples, (keep existing compat) (#226) (@github-actions[bot])
• CompatHelper: bump compat for MLUtils to 0.4 for package ImageNet, (keep existing compat) (#227) (@github-actions[bot])
• CompatHelper: bump compat for MLUtils to 0.4 for package DDIM, (keep existing compat) (#228) (@github-actions[bot])
• Functor ambiguity fix (#229) (@avik-pal)
• Add all compats together (#238) (@avik-pal)
• CompatHelper: bump compat for Turing to 0.24 for package examples, (keep existing compat) (#241) (@github-actions[bot])
• CompatHelper: bump compat for JET to 0.7 for package test, (keep existing compat) (#251) (@github-actions[bot])
• [WIP] Use Extensions for Flux2Lux (#261) (@avik-pal)
• Cleaner test workflow (#262) (@avik-pal)
• Add a patch for #243 (#263) (@avik-pal)
• Update LuxLib dependencies (#265) (@avik-pal)
• Dropping Julia 1.6 support for Lux (#266) (@avik-pal)
• Purge unnecessary dependencies into weak dependencies (#267) (@avik-pal)
• Add ForwardDiff Extension: Dropout (#269) (@avik-pal)
• Add Tracker as an Extension (#272) (@avik-pal)
• CompatHelper: bump compat for AbstractDifferentiation to 0.5 for package examples, (keep existing compat) (#273) (@github-actions[bot])
• Some Improvements (#274) (@avik-pal)
• Tracker has some of the rules (#275) (@avik-pal)
• Temporary CA + Tracker Patches (#276) (@avik-pal)
• Add CUDA and AMDGPU trigger packages (#277) (@avik-pal)
• ReverseDiff Extension (#280) (@avik-pal)
• Bump peter-evans/create-pull-request from 3 to 4 (#283) (@dependabot[bot])
• Bump actions/cache from 1 to 3 (#284) (@dependabot[bot])
• Bump actions/checkout from 1 to 3 (#285) (@dependabot[bot])
• Return the history for Recurrence (#287) (@avik-pal)
• Truncate tuples and namedtuples (#290) (@avik-pal)
• [WIP] Remove projects from lib to LuxDL (#291) (@avik-pal)
• Patch freeze (#292) (@avik-pal)
• Add dispatch for no activation (#293) (@avik-pal)
• Remove weakdeps from deps (#295) (@avik-pal)
• Try restoring lts support (#296) (@avik-pal)
• Testing using LuxTestUtils.jl (#297) (@avik-pal)
• CompatHelper: bump compat for Boltz to 0.2 for package ImageNet, (kee… (#298) (@avik-pal)
• Bump peter-evans/create-pull-request from 4 to 5 (#299) (@dependabot[bot])
• remove Dataloaders (#300) (@avik-pal)
• Update docs (#301) (@avik-pal)
• Fix bug in recurrence ordering (#303) (@avik-pal)
• Update LuxComponentArraysExt.jl (#304) (@avik-pal)
• CompatHelper: bump compat for Turing to 0.25 for package examples, (keep existing compat) (...

LuxLib-v1.4.1

LuxLib LuxLib-v1.4.1

Diff since LuxLib-v1.4.0

Merged pull requests:

• feat: update ConvMixer to support reactant (#1063) (@avik-pal)
• fix: update default rng for reactant (#1152) (@avik-pal)
• Change update step to thousands in PINN 2D PDE (#1153) (@abhro)
• CompatHelper: add new compat entry for ProgressTables at version 0.1 for package CIFAR10, (keep existing compat) (#1154) (@github-actions[bot])
• device(NN) should only give warning once (#1156) (@vpuri3)
• feat: conditional VAE (#1157) (@avik-pal)
• fix: ConditionalVAE on CI (#1159) (@avik-pal)
• CompatHelper: add new compat entry for Optimisers at version 0.4 for package ConditionalVAE, (keep existing compat) (#1161) (@github-actions[bot])
• CompatHelper: add new compat entry for StableRNGs at version 1 for package ConditionalVAE, (keep existing compat) (#1162) (@github-actions[bot])
• CompatHelper: add new compat entry for Comonicon at version 1 for package ConditionalVAE, (keep existing compat) (#1163) (@github-actions[bot])
• [MLDataDevices] Bump deps (#1164) (@pxl-th)
• docs: migrate most examples to Reactant (#1180) (@avik-pal)
• chore: bump crate-ci/typos from 1.28.4 to 1.29.4 (#1183) (@dependabot[bot])
• fix: pass in RNG to shuffle (#1188) (@avik-pal)

Closed issues:

• Random Numbers & Reactant (#1131)
• Recurrent cells cannot be chained with other layers (#1155)
• sparse_init doesn't use provided rng fully (#1185)
• Incorrect IR generated for some neural networks (#1186)
• WeightInitializers.DeviceAgnostic doesn't respect Reactant (#1187)

LuxCore-v1.2.2

LuxCore LuxCore-v1.2.2

Diff since LuxCore-v1.2.1

Merged pull requests:

• feat: update ConvMixer to support reactant (#1063) (@avik-pal)
• test: re-enable flux testing (#1123) (@avik-pal)
• chore: bump minimum Reactant version (#1125) (@avik-pal)
• fix: try fixing cuda install in tests (#1126) (@avik-pal)
• docs: run partial dataset only on CI (#1128) (@avik-pal)
• chore: bump crate-ci/typos from 1.28.1 to 1.28.2 (#1130) (@dependabot[bot])
• fix: preserve object when device is same (#1133) (@avik-pal)
• fix: use functors for testing wrapped arrays (#1134) (@avik-pal)
• fix: remove old patches around reactant bug (#1135) (@avik-pal)
• CompatHelper: bump compat for Flux in [weakdeps] to 0.16, (keep existing compat) (#1136) (@github-actions[bot])
• chore: bump crate-ci/typos from 1.28.2 to 1.28.3 (#1137) (@dependabot[bot])
• fix: update to new reactant changes (#1140) (@avik-pal)
• chore: bump crate-ci/typos from 1.28.3 to 1.28.4 (#1144) (@dependabot[bot])
• don't declare implicitly exported functions public (#1147) (@simeonschaub)
• use return_type instead of _return_type (#1148) (@simeonschaub)
• feat: more nested AD rules (#1151) (@avik-pal)
• fix: update default rng for reactant (#1152) (@avik-pal)
• Change update step to thousands in PINN 2D PDE (#1153) (@abhro)
• CompatHelper: add new compat entry for ProgressTables at version 0.1 for package CIFAR10, (keep existing compat) (#1154) (@github-actions[bot])
• device(NN) should only give warning once (#1156) (@vpuri3)
• feat: conditional VAE (#1157) (@avik-pal)
• fix: ConditionalVAE on CI (#1159) (@avik-pal)
• CompatHelper: add new compat entry for Optimisers at version 0.4 for package ConditionalVAE, (keep existing compat) (#1161) (@github-actions[bot])
• CompatHelper: add new compat entry for StableRNGs at version 1 for package ConditionalVAE, (keep existing compat) (#1162) (@github-actions[bot])
• CompatHelper: add new compat entry for Comonicon at version 1 for package ConditionalVAE, (keep existing compat) (#1163) (@github-actions[bot])
• [MLDataDevices] Bump deps (#1164) (@pxl-th)
• docs: migrate most examples to Reactant (#1180) (@avik-pal)
• chore: bump crate-ci/typos from 1.28.4 to 1.29.4 (#1183) (@dependabot[bot])
• fix: pass in RNG to shuffle (#1188) (@avik-pal)

Closed issues:

• Immutable Arrays (#8)
• Downstream Compat Updates (#880)
• Re-enable Flux compatibility testing (#1070)
• Documentation Build Stalls (#1120)
• CUDA Test CI is broken (#1121)
• [MLDataDevices] devices don't preserve identity (#1129)
• Random Numbers & Reactant (#1131)
• Problem with Lux & SymbolicsLuxExt (#1132)
• How to implement a detach operation similar to Pytorch? (#1138)
• Unexpected handling of LR Schedulers in TrainState (#1143)
• Directly construct Optimiser state on Reactant buffers (#1145)
• (AbstractDevice)(x) should respect Adapt.adapt_structure (#1149)
• CUDA 2nd order AD with MaxPool and logsoftmax (#1150)
• Recurrent cells cannot be chained with other layers (#1155)
• sparse_init doesn't use provided rng fully (#1185)
• Incorrect IR generated for some neural networks (#1186)
• WeightInitializers.DeviceAgnostic doesn't respect Reactant (#1187)

MLDataDevices-v1.6.7

MLDataDevices MLDataDevices-v1.6.7

Diff since MLDataDevices-v1.6.6

Merged pull requests:

• feat: update ConvMixer to support reactant (#1063) (@avik-pal)
• Change update step to thousands in PINN 2D PDE (#1153) (@abhro)
• CompatHelper: add new compat entry for ProgressTables at version 0.1 for package CIFAR10, (keep existing compat) (#1154) (@github-actions[bot])
• device(NN) should only give warning once (#1156) (@vpuri3)
• feat: conditional VAE (#1157) (@avik-pal)
• fix: ConditionalVAE on CI (#1159) (@avik-pal)
• CompatHelper: add new compat entry for Optimisers at version 0.4 for package ConditionalVAE, (keep existing compat) (#1161) (@github-actions[bot])
• CompatHelper: add new compat entry for StableRNGs at version 1 for package ConditionalVAE, (keep existing compat) (#1162) (@github-actions[bot])
• CompatHelper: add new compat entry for Comonicon at version 1 for package ConditionalVAE, (keep existing compat) (#1163) (@github-actions[bot])
• [MLDataDevices] Bump deps (#1164) (@pxl-th)

Closed issues:

• Recurrent cells cannot be chained with other layers (#1155)

v1.4.4

Lux v1.4.4

Diff since v1.4.3

Merged pull requests:

• feat: update ConvMixer to support reactant (#1063) (@avik-pal)
• feat: more nested AD rules (#1151) (@avik-pal)
• fix: update default rng for reactant (#1152) (@avik-pal)
• Change update step to thousands in PINN 2D PDE (#1153) (@abhro)

Closed issues:

• Random Numbers & Reactant (#1131)
• CUDA 2nd order AD with MaxPool and logsoftmax (#1150)

MLDataDevices-v1.6.6

MLDataDevices MLDataDevices-v1.6.6

Diff since MLDataDevices-v1.6.5

Merged pull requests:

• fix: remove old patches around reactant bug (#1135) (@avik-pal)
• CompatHelper: bump compat for Flux in [weakdeps] to 0.16, (keep existing compat) (#1136) (@github-actions[bot])
• chore: bump crate-ci/typos from 1.28.2 to 1.28.3 (#1137) (@dependabot[bot])
• fix: update to new reactant changes (#1140) (@avik-pal)
• chore: bump crate-ci/typos from 1.28.3 to 1.28.4 (#1144) (@dependabot[bot])
• don't declare implicitly exported functions public (#1147) (@simeonschaub)
• use return_type instead of _return_type (#1148) (@simeonschaub)
• feat: more nested AD rules (#1151) (@avik-pal)
• fix: update default rng for reactant (#1152) (@avik-pal)

Closed issues:

• Random Numbers & Reactant (#1131)
• Problem with Lux & SymbolicsLuxExt (#1132)
• How to implement a detach operation similar to Pytorch? (#1138)
• Unexpected handling of LR Schedulers in TrainState (#1143)
• Directly construct Optimiser state on Reactant buffers (#1145)
• (AbstractDevice)(x) should respect Adapt.adapt_structure (#1149)
• CUDA 2nd order AD with MaxPool and logsoftmax (#1150)

LuxLib-v1.4.0

LuxLib LuxLib-v1.4.0

Diff since LuxLib-v1.3.11

Merged pull requests:

• feat: more nested AD rules (#1151) (@avik-pal)

Closed issues:

• CUDA 2nd order AD with MaxPool and logsoftmax (#1150)

v1.4.3

Lux v1.4.3

Diff since v1.4.2

Merged pull requests:

• CompatHelper: bump compat for Flux in [weakdeps] to 0.16, (keep existing compat) (#1136) (@github-actions[bot])
• chore: bump crate-ci/typos from 1.28.2 to 1.28.3 (#1137) (@dependabot[bot])
• fix: update to new reactant changes (#1140) (@avik-pal)
• chore: bump crate-ci/typos from 1.28.3 to 1.28.4 (#1144) (@dependabot[bot])
• don't declare implicitly exported functions public (#1147) (@simeonschaub)
• use return_type instead of _return_type (#1148) (@simeonschaub)

Closed issues:

• Problem with Lux & SymbolicsLuxExt (#1132)
• How to implement a detach operation similar to Pytorch? (#1138)
• Unexpected handling of LR Schedulers in TrainState (#1143)
• Directly construct Optimiser state on Reactant buffers (#1145)
• (AbstractDevice)(x) should respect Adapt.adapt_structure (#1149)