v0.6.22

Zygote v0.6.22

Diff since v0.6.21

Closed issues:

• Complex gradient on real function with complex intermediates (#342)
• Adjoints for functions of specialized matrices (#402)
• Gradient of inv and logdet UpperTriangular matrix is not UpperTriangular (#431)
• Trying to differentiate a function with complex and real inputs results in error (#917)

Merged pull requests:

• Use ProjectTo in broadcasting & gradient (#1044) (@mcabbott)
• Allow buffer typevar in CuArray type (#1062) (@DhairyaLGandhi)

v0.6.21

Zygote v0.6.21

Diff since v0.6.20

Closed issues:

• Relation to capstan and cassette.jl (#1)
• Error when taking gradient involving zeros(eltype(x), ...). (#72)
• Error when taking gradient involving a range broadcast (#73)
• Deprecation warning in tests (#76)
• Error when trying to define gradient for addition (#84)
• CUDA broadcast over * and + blows up taking gradient (#113)
• cholesky keywords (#121)
• Division causes f32 -> f64 conversion in backward pass (#136)
• differentiating over x isa NTuple{<:Any, <:Integer} causes julia crush (#158)
• simple MLP gradient fails when dropout is included (#171)
• BoundsError in mutable array branch (#172)
• Trouble with reshape() (#187)
• simple broadcast not working (#195)
• Lazy adjoint/transpose gradient fails (#213)
• getindex for Adjoint does not back-propagate (#217)
• Add support for logpdf of Poisson (#281)
• Gradients for cumsum and cumprod (#282)
• Support for tuple and function argument destructuring (#303)
• Cannot differentiate through dictionary (#817)
• Scalar indexing error from GPU matmul against Zygote.OneElement (#1005)
• documentation seems stale because of CI failure (#1021)
• Typo in docs? (#1056)
• Second derivatives not working (#1067)

Merged pull requests:

• Support functions that splat namedtuples as keyword arguments (#1059) (@oxinabox)
• fix pair getfield pullback (#1063) (@CarloLucibello)
• Don't use dual numbers in broadcasting within 2nd derivatives (#1068) (@mcabbott)
• Add short-circuit to rrule_via_ad (#1069) (@mcabbott)

v0.6.20

Zygote v0.6.20

Diff since v0.6.19

Closed issues:

• Batched RNN on GPU: CuArray only supports bits types (#814)
• gradient doesn't seem to work with Julia1.6 (#818)

Merged pull requests:

• Fix docstring cross-reference (#1054) (@st--)
• Support kwargs in rrule_via_ad (#1055) (@mzgubic)

v0.6.19

Zygote v0.6.19

Diff since v0.6.18

v0.6.18

Zygote v0.6.18

Diff since v0.6.17

Closed issues:

• Error (no method matching iterate(::Nothing)) when counting non-zero elements in loss function (#1027)
• Fix ChainRules deprecation warnings for InplaceableThunk (#1030)
• Can't construct Params with empty parameters (#1032)
• julia examples/mnist_mlp.jl -> ERROR: LoadError: UndefVarError: data not defined (#1046)
• Zygote+RandomNumbers Usage Broken: "ArgumentError: type does not have a definite number of fields" (#1048)

Merged pull requests:

• Tidy up a few things in broadcasting (#1018) (@mcabbott)
• Move to ChainRulesCore v1.0 (in OptingOut) (#1035) (@oxinabox)
• Fix UndefVarError (#1038) (@dlfivefifty)
• Extra fixes for ChainRulesCore @1.0 (#1039) (@mzgubic)
• Update CompatHelper script (#1040) (@DhairyaLGandhi)
• remove artifact warning (#1047) (@mzgubic)
• Remove examples folder (#1050) (@mcabbott)
• Bump patch (#1052) (@willtebbutt)

v0.6.17

Zygote v0.6.17

Diff since v0.6.16

Closed issues:

• Base.Fix1 and Base.Fix2 are much slower than closures (#957)
• It seems a breaking change was introduced after v0.6.14 (#1029)

Merged pull requests:

• construct Params with empty tuple (#1031) (@CarloLucibello)

v0.6.16

Zygote v0.6.16

Diff since v0.6.15

Merged pull requests:

• Fix resolver message at initialization (#1028) (@DhairyaLGandhi)

v0.6.15

Zygote v0.6.15

Diff since v0.6.14

Closed issues:

• custom getproperty of mutable struct fails (#685)
• Regression from v0.6.3 to v0.6.8: Mutating arrays error when not mutating anything (#941)
• Error when taking gradient of functions using similar with CUDA CuArray (#997)
• gradient is slow when a lambda expression is passed, due to 99% compilation time of the lambda expression. (#1007)
• Use ismutabletype on 1.7 (#1023)

Merged pull requests:

• fix broken link (#878) (@ho-oto)
• Add scalar method for vcat with number (#880) (@DhairyaLGandhi)
• Improve performance of Base.Fix1 and Base.Fix2 (#958) (@devmotion)
• Differentiate push! with implicit Params (#992) (@DhairyaLGandhi)
• Faster generic broadcasting (#1001) (@mcabbott)
• use rrules even when all the arguments are types (#1006) (@oxinabox)
• Use abstract GPU types for broadcast. (#1009) (@akashkgarg)
• Add convenience function to return primal & cotangents together (#1010) (@mcabbott)
• Documentation edit on implicit parameters (#1016) (@JordiBolibar)
• Whitespace fixes (#1019) (@DhairyaLGandhi)
• CompatHelper: bump compat for "FillArrays" to "0.12" (#1020) (@github-actions[bot])
• documentation updates (#1022) (@aviatesk)
• Make broadcasting over Params in the gradient an error (#1024) (@mcabbott)
• make sure to throw explicit CompileError (#1026) (@aviatesk)

v0.6.14

Zygote v0.6.14

Diff since v0.6.13

Merged pull requests:

• Make interface2.jl code around generating pullbacks via decomposition (#1002) (@oxinabox)
• use Base.ismutabletype (#1003) (@mcabbott)

v0.6.13

Zygote v0.6.13

Diff since v0.6.12

Closed issues:

• Move over adjoints from Flux (#81)
• Two errors on Julia 1.6 (#897)
• Wrong Gradient with dot product of broadcast (#916)
• Tangent not defined (#980)
• Error when taking gradient of scalar function using rand.() (#996)

Merged pull requests:

• delete rule for prod to use ChainRules' (#988) (@oxinabox)
• Use ChainRules RuleConfig (#990) (@oxinabox)
• fix #996 (#998) (@simeonschaub)