v0.4.13

Zygote v0.4.13

Diff since v0.4.12

Closed issues:

• Logging and String Interpolation Errors in gradient (#560)
• Array Padding (Plain Julia/Images.jl/PaddedViews.jl) (#561)
• Gradients of broadcasted exp broken with CuArrays v2.0.1 (#567)
• Compat with ArrayLayout 0.2 (#576)

Merged pull requests:

• AbstractFFTs, rfft, plans, and a couple of tests (#270) (@dsweber2)
• nograd string, repr (#562) (@CarloLucibello)
• Fix broadcast for CuArray with CuArrays 2.0 (#565) (@ararslan)
• fix gpu integration tests (#569) (@CarloLucibello)
• add clarifying brackets to docs on what pullback does (#573) (@oxinabox)
• Fix GPU broadcasting tests (#574) (@ararslan)
• CompatHelper: bump compat for "ArrayLayouts" to "0.2" (#579) (@github-actions[bot])
• CompatHelper: add new compat entry for "AbstractFFTs" at version "0.5" (#580) (@github-actions[bot])
• Make Cholesky adjoint GPU-friendly (#581) (@aterenin)

v0.4.12

Zygote v0.4.12

Diff since v0.4.11

Closed issues:

• efficient gradient for fft (#204)
• WARNING: Method definition _pullback (#554)

Merged pull requests:

• Remove extra print / println @nograds (#555) (@willtebbutt)

v0.4.11

Zygote v0.4.11

Diff since v0.4.10

Closed issues:

• RFC: return full gradient of all arguments in gradient (#535)
• adjoint function for complex functions (#540)

Merged pull requests:

• Restrict some linear algebra adjoints to matrices (#545) (@mohamed82008)

v0.4.10

Zygote v0.4.10

Diff since v0.4.9

Closed issues:

• Three ways to fail taking gradients in sum() (#231)
• sum fails to capture closure variables (#314)
• vcat with literal fails (#440)
• vcat error for appending scalars (#538)

Merged pull requests:

• Adds a nograd function to encapsulate an operation without gradient (#465) (@cossio)
• CompatHelper: add new compat entry for "ArrayLayouts" at version "0.1" (#533) (@github-actions[bot])
• fix adjoint sum(f, xs) (#534) (@CarloLucibello)
• Add a check kwarg to cholesky adjoint and bug fix (#539) (@mohamed82008)
• vcat vectors + numbers (#542) (@mcabbott)

v0.4.9

Zygote v0.4.9

Diff since v0.4.8

Closed issues:

• vcat of a vector with a number is not differentiable (#516)
• tag new release (#528)

Merged pull requests:

• fix a few adjoints for complex arrays (#519) (@marius311)
• tag new release (#527) (@CarloLucibello)
• Simplify global variable handling (#529) (@MikeInnes)
• Gradient for batched_mul (#531) (@mcabbott)
• fix conv on cpu with fillarays (#532) (@CarloLucibello)

v0.4.8

Zygote v0.4.8

Diff since v0.4.7

Closed issues:

• Adjoint of literal_pow fails when grad is calculated on negative input (#247)
• softmax adjoint does not support the dims keyword (#425)
• Tag? (#427)
• gradient on Distances.jl evaluate causes Julia to crash (#429)
• Readme Error (#470)
• Need an adjoint for constructor Base.OneTo{Int64} (#483)
• Julia Crash (#486)
• Zygote suffers poor performance w/ type conversions (#491)
• Adjoint for filter (#492)
• Cannot calculate gradient. (#511)
• Literal arrays are not differentiable (#513)

Merged pull requests:

• Simplest prod(x; dims) gradient (#112) (@mcabbott)
• adjoint for cumsum (#284) (@mcabbott)
• Fix Euclidean and SqEuclidean distances (#463) (@devmotion)
• use IdDict in grad_mut fallback for IdDict (#474) (@ChrisRackauckas)
• Add @Nograd for axes (#485) (@AzamatB)
• Fix forwarddiff docstring (#493) (@ChrisRackauckas)
• Gradient for collect (#498) (@mcabbott)
• Fix a bug in ∇getindex (#499) (@mcabbott)
• Gradient for filter (#500) (@mcabbott)
• delete! for Params (#505) (@CarloLucibello)
• Install TagBot as a GitHub Action (#507) (@JuliaTagBot)
• Add adjoint for repeat(x::AbstractVecOrMat, m, n) (#508) (@AzamatB)
• nograd <= (#509) (@matsueushi)
• Gradient for power (#512) (@mcabbott)
• Allow non-uniform Number types in hvcat (#514) (@mcabbott)
• Fix Flux.flip by providing an adjoint for Base.reverse (#515) (@tanhevg)
• Remove get! macro (#517) (@matsueushi)
• Special adjoint for broadcasted literal pow (#518) (@haampie)
• Correct minor typos. (#520) (@ludgerpaehler)
• fix prod with tuple arg (#523) (@CarloLucibello)
• getindex for Float32[x,y] etc. (#524) (@mcabbott)
• Adjoints for StepRange (#525) (@hessammehr)

v0.4.7

v0.4.7 (2020-01-29)

Diff since v0.4.6

Closed issues:

• Readme outdated (#484)
• Gradient of exp(real matrix) should be real (#475)
• Intermediate complex number causes complex gradient (#472)
• Converting UniformScaling to Matrix causes mutation (#468)

Merged pull requests:

• Make version condition on Core._apply_iterate precise (#480) (andreasnoack)
• make docs build (#479) (oxinabox)
• Fix I - A (#477) (mohamed82008)
• make Buffer iterable and use it for Params (#476) (piever)
• Adjoint for sort (#473) (MikeInnes)
• Fix converting UniformScaling to Matrix (#469) (mohamed82008)
• Include rank in conv and pooling gradcheck testset names (#461) (oxinabox)
• matrix exp(A) returns real-valued adjoint for real-valued A (#409) (sdewaele)

v0.4.6

v0.4.6 (2020-01-14)

Diff since v0.4.5

Merged pull requests:

• Update .travis.yml (#460) (MikeInnes)
• Limit maximum and minimum adjoints to arrays (#459) (mohamed82008)
• Fix gitlab CI (#458) (MikeInnes)
• Fix Fill ctor (#445) (willtebbutt)
• Use dispatch for accum nothing cases (#444) (oxinabox)
• Add CuArray convert adjoint (#423) (dhairyagandhi96)

v0.4.5

v0.4.5 (2020-01-10)

Diff since v0.4.4

Closed issues:

• Zygote fails when an adjoint hits the GC (#451)
• Basic function errors (#426)
• ERROR: InitError: LoadError: Can't differentiate loopinfo expression (#393)

Merged pull requests:

• 1.4 updates (#453) (MikeInnes)
• Handle dims keyword argument in softmax adjoint (#443) (tanhevg)

v0.4.4

v0.4.4 (2020-01-08)

Diff since v0.4.3

Merged pull requests:

• support for mutable globals (#448) (MikeInnes)
• Push for buffers (#447) (MikeInnes)
• Add adjoints for UnitUpperTriangular. (#442) (aterenin)
• Fix example result in adjoints.md (#439) (cdsousa)