Releases: FluxML/Zygote.jl
Releases · FluxML/Zygote.jl
v0.6.42
Zygote v0.6.42
Closed issues:
- Flux.train error when using reduce hcat in loss (#626)
- Broadcasting fails for functions with adjoint 'nothing' (#642)
- Incorrect Wirtinger for
prod
operation (#744) - Surprising error when constructing diagonal matrix (#771)
- dropdims gives MethodError in v0.5.6 when training with Flux.jl (#784)
- accum NamedTuple sometimes silently drops derivatives (#861)
- Dot product differentiation (#982)
unique()
error in pullback (#1053)- Adjoint of
cholesky
is hard-coded for the CPU (#1210) - Deprecate
@nograd
anddropgrad
in favour ofChainRulesCore.@non_differentiable
andChainRulesCore.ignore_derivatives
(#1235) - Remove
@adjoint
foradjoint
/Adjoint
in favor ofChainRules
(#1257) - Error message link 404 error (#1260)
Merged pull requests:
v0.6.41
Zygote v0.6.41
Closed issues:
- Limitations Docs (#1077)
- DiffRules warning (#1155)
- Doctests running twice (#1199)
vcat(::Vector, ::Pair)
gradient bug (#1223)- Zygote.gradient returns
nothing
instead of `NotImplemented (#1227) - Rules for mutating functions, @adjoint! and its documentation (#1228)
- Possible mistake in complex tests (#1229)
- Possible mistake in complex tests (#1230)
- Gradient missing when using Params within Dicts (#1241)
- nested gradient fails with "Mutating arrays is not supported" (#1244)
- Incorrect gradient when using Flux.params (#1251)
- Gradients for nested struct fields are super slow (#1256)
Merged pull requests:
- remove
@adjoint function cholesky
(#1114) (@st--) - Fix deprecations in DiffRules 1.4, add forward-mode rules for LogExpFunctions (#1172) (@devmotion)
- Make LogExpFunctions a proper dependency and clean adjoints (#1221) (@devmotion)
- Add warnings to ForwardDiff functions (#1224) (@mcabbott)
- Use CR adjoint for
logdet(::Cholesky)
(#1226) (@devmotion) - Fix non-holomorphic tests (#1231) (@DomCRose)
- Fix #1241 (#1242) (@bogiebro)
- deprecate
dropgrad
andignore
(#1245) (@mzgubic) - Make errors for array mutation, try/catch, and foreign call nicer (+ docs) (#1246) (@darsnack)
- rm rules for
maximum
,minimum
,dropdims
(#1250) (@mcabbott) - rm rules for
eachslice
,cumsum
(#1253) (@mcabbott) - Run doctests only once (#1255) (@Saransh-cpp)
v0.6.40
Zygote v0.6.40
Closed issues:
- Function mod2pi + list of implemented functions (#118)
- Unexpected error with using the
normalize
from LinearAlgebra (#1190) - Gradient Dimension Mismatch (#1192)
Merged pull requests:
- Make chunk threshold configurable (#1213) (@lassepe)
- Use
setglobal!
on nightly (#1219) (@ToucheSir) - Remove unnecessary generated functions (#1220) (@MasonProtter)
v0.6.39
Zygote v0.6.39
Merged pull requests:
- Actually correct
Base.in(x, ps::Params)
(#1207) (@ToucheSir) - Remove
cat
adjoint in favour of ChainRules (#1212) (@ToucheSir) - Correct spelling mistake (#1214) (@zsoerenm)
v0.6.38
Zygote v0.6.38
Closed issues:
- non-differentiable error for code that needn't be differentiable (#175)
- Performance of gradient depends on size of implicitly used array (#1180)
- Initializing buffer twice in loop results in a corrupted array (#1191)
- Mismatch gradient output with compared to the docs. (#1193)
- Params iteration is type unstable (#1202)
- AD error due to interaction of missing gradient of
besselk
and struct with vector (#1204)
Merged pull requests:
- Ton of doctests added (#1194) (@arcAman07)
- Fixing spelling error in the docs (#1197) (@arcAman07)
- Fix type stability of Params.order (#1203) (@ancapdev)
- Handle
ChainRulesCore.NotImplemented
(#1205) (@devmotion)
v0.6.37
Zygote v0.6.37
Merged pull requests:
- Use Base.IdSet (#1183) (@ToucheSir)
- Fixing type ambiguity of
unbroadcast
(#1185) (@SebastianAment)
v0.6.36
v0.6.35
Zygote v0.6.35
Closed issues:
- Matrix (and higher) comprehensions causes stackoverflow (#176)
- ERROR: LoadError: Need an adjoint for constructor SparseMatrixCSC{Float64,Int64}. Gradient is of type Array{Float64,2}. (#742)
- gradient calculation with explicit type cast is broken (#810)
- change repo description from "Intimate Affection Auditor" (#816)
- README should not mention
@adjoint
(#1105) - Significant compile time latency in Flux with default optimization (#1126)
- Differentiation of a scalar times array times array results in non-differentiable expression (#1140)
- Error in
gradient
of customAbstractMatrix
subtypes (#1146) - Non-deterministic
AssertionError
when differentiatingThreadsX.sum
(#1159) - Insane precompile times in Metalhead models (#1160)
- Type error with
zip
and heterogeneous vectors in pullback (#1162) - Unable to differentiate through Einstein sums from Einsum.jl -- Mutating arrays is not supported (#1164)
Merged pull requests:
- Add codecov badge (#1151) (@ToucheSir)
- Update codecov action and only run on stable linux (#1152) (@ToucheSir)
ntuple
for_restore
regardless of length (#1163) (@SBuercklin)- CompatHelper: bump compat for "FillArrays" to "0.13" (#1165) (@github-actions[bot])
- CompatHelper: bump compat for "NaNMath" to "1" (#1166) (@github-actions[bot])
- update README to reference ChainRulesCore to define custom gradients (#1169) (@mzgubic)
- Broadcast rule for types (#1171) (@mcabbott)
v0.6.34
Zygote v0.6.34
Closed issues:
Merged pull requests:
- Fix Buffer when storing arrays (#1134) (@avik-pal)
- Make value returned by
copy!
consistent (#1136) (@pakk-minidose) - Update Buildkite config for 1.6 LTS and 1.7 (#1142) (@ToucheSir)
- Downstream test for Molly.jl (#1145) (@jgreener64)
v0.6.33
Zygote v0.6.33
Closed issues:
- Support sparse arrays/matrices as output? (#163)
- AD through cuFFT fails, (type instability / broadcast) (#1129)
- Broadcast gradient error (#1132)
Merged pull requests:
- CompatHelper: bump compat for "SpecialFunctions" to "2" (#1122) (@github-actions[bot])
- Fix incorrect
@forward
ing ofBase.in
onParams
(#1130) (@ToucheSir) - Tag for SpecialFunctions@2 (#1138) (@DhairyaLGandhi)