Merge #1238

1238: Fix inline code block r=dhairyagandhi96 a=harryscholes

### PR Checklist

- [ ] Tests are added
- [ ] Entry in NEWS.md
- [x] Documentation, if applicable
- [ ] Final review from `@MikeInnes` or `@dhairyagandhi96` (for API changes).


Co-authored-by: harryscholes <harryscholes@gmail.com>

.github/pull_request_template.md

@@ -0,0 +1,12 @@
+[Please delete this text and describe your change here.
+For bugfixes, please detail the bug and include a test case which your patch fixes.
+If you are adding a new feature, please clearly describe the design, its rationale, the possible alternatives considered.
+It is easiest to merge new features when there is clear precedent in other systems; we need to know we're taking
+the right direction since it can be hard to change later.]
+
+### PR Checklist
+
+- [ ] Tests are added
+- [ ] Entry in NEWS.md
+- [ ] Documentation, if applicable
+- [ ] Final review from `@MikeInnes` or `@dhairyagandhi96` (for API changes).

Manifest.toml

@@ -8,35 +8,35 @@ version = "0.5.0"
 
 [[AbstractTrees]]
 deps = ["Markdown"]
-git-tree-sha1 = "86d092c2599f1f7bb01668bf8eb3412f98d61e47"
+git-tree-sha1 = "33e450545eaf7699da1a6e755f9ea65f14077a45"
 uuid = "1520ce14-60c1-5f80-bbc7-55ef81b5835c"
-version = "0.3.2"
+version = "0.3.3"
 
 [[Adapt]]
 deps = ["LinearAlgebra"]
-git-tree-sha1 = "c88cfc7f9c1f9f8633cddf0b56e86302b70f64c5"
+git-tree-sha1 = "fd04049c7dd78cfef0b06cdc1f0f181467655712"
 uuid = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
-version = "1.0.1"
+version = "1.1.0"
 
 [[ArrayLayouts]]
 deps = ["FillArrays", "LinearAlgebra"]
-git-tree-sha1 = "41956a49a8a4fefa1bf6664bca4a3035aba4c3a0"
+git-tree-sha1 = "a504dca2ac7eda8761c8f7c1ed52427a1be75a3c"
 uuid = "4c555306-a7a7-4459-81d9-ec55ddd5c99a"
-version = "0.2.3"
+version = "0.2.6"
 
 [[Base64]]
 uuid = "2a0f44e3-6c83-55bd-87e4-b1978d98bd5f"
 
 [[BinaryProvider]]
-deps = ["Libdl", "SHA"]
+deps = ["Libdl", "Logging", "SHA"]
-git-tree-sha1 = "5b08ed6036d9d3f0ee6369410b830f8873d4024c"
+git-tree-sha1 = "ecdec412a9abc8db54c0efc5548c64dfce072058"
 uuid = "b99e7846-7c00-51b0-8f62-c81ae34c0232"
-version = "0.5.8"
+version = "0.5.10"
 
 [[CEnum]]
-git-tree-sha1 = "62847acab40e6855a9b5905ccb99c2b5cf6b3ebb"
+git-tree-sha1 = "1b77a77c3b28e0b3f413f7567c9bb8dd9bdccd14"
 uuid = "fa961155-64e5-5f13-b03f-caf6b980ea82"
-version = "0.2.0"
+version = "0.3.0"
 
 [[CUDAapi]]
 deps = ["Libdl", "Logging"]
@@ -46,21 +46,21 @@ version = "4.0.0"
 
 [[CUDAdrv]]
 deps = ["CEnum", "CUDAapi", "Printf"]
-git-tree-sha1 = "e650cbaee92b60433313157926b1e80d0c3a0e2e"
+git-tree-sha1 = "f56bbf18c86bcff7a961a32a4947a5abb2963a29"
 uuid = "c5f51814-7f29-56b8-a69c-e4d8f6be1fde"
-version = "6.2.2"
+version = "6.3.0"
 
 [[CUDAnative]]
-deps = ["Adapt", "BinaryProvider", "CEnum", "CUDAapi", "CUDAdrv", "Cthulhu", "DataStructures", "InteractiveUtils", "LLVM", "Libdl", "MacroTools", "Pkg", "Printf", "TimerOutputs"]
+deps = ["Adapt", "BinaryProvider", "CEnum", "CUDAapi", "CUDAdrv", "ExprTools", "GPUCompiler", "LLVM", "Libdl", "Pkg", "Printf"]
-git-tree-sha1 = "d1fc99635d0002c8a819b78cb1f441eb44310725"
+git-tree-sha1 = "ac86db2b05fdfec96b011e25a504ffe7476e8a68"
 uuid = "be33ccc6-a3ff-5ff2-a52e-74243cff1e17"
-version = "3.0.2"
+version = "3.1.0"
 
 [[CodeTracking]]
 deps = ["InteractiveUtils", "UUIDs"]
-git-tree-sha1 = "0becdab7e6fbbcb7b88d8de5b72e5bb2f28239f3"
+git-tree-sha1 = "cab4da992adc0a64f63fa30d2db2fd8bec40cab4"
 uuid = "da1fd8a2-8d9e-5ec2-8556-3022fb5608a2"
-version = "0.5.8"
+version = "0.5.11"
 
 [[CodecZlib]]
 deps = ["TranscodingStreams", "Zlib_jll"]
@@ -70,15 +70,15 @@ version = "0.7.0"
 
 [[ColorTypes]]
 deps = ["FixedPointNumbers", "Random"]
-git-tree-sha1 = "c4c1cca28748906265ed62c788d6fe6f0134d264"
+git-tree-sha1 = "c73d9cfc2a9d8433dc77f5bff4bddf46b1d78c20"
 uuid = "3da002f7-5984-5a60-b8a6-cbb66c0b333f"
-version = "0.10.0"
+version = "0.10.3"
 
 [[Colors]]
 deps = ["ColorTypes", "FixedPointNumbers", "InteractiveUtils", "Reexport"]
-git-tree-sha1 = "2fdeb981ebcf52cd800ddb6a0aa5eac34153552d"
+git-tree-sha1 = "1e9bba7984e78aa8cdeea7f9f7cc984ad4e4b1c7"
 uuid = "5ae59095-9a9b-59fe-a467-6f913c188581"
-version = "0.12.0"
+version = "0.12.2"
 
 [[CommonSubexpressions]]
 deps = ["Test"]
@@ -93,27 +93,27 @@ uuid = "e66e0078-7015-5450-92f7-15fbd957f2ae"
 version = "0.3.3+0"
 
 [[Cthulhu]]
-deps = ["CodeTracking", "InteractiveUtils", "REPL", "Unicode"]
+deps = ["CodeTracking", "InteractiveUtils", "REPL", "UUIDs", "Unicode"]
-git-tree-sha1 = "484790098c85c26f8e59051f8ff1a0745c034a7d"
+git-tree-sha1 = "f3643e78353199d3097821e806348bd83f364155"
 uuid = "f68482b8-f384-11e8-15f7-abe071a5a75f"
-version = "1.0.1"
+version = "1.1.1"
 
 [[CuArrays]]
 deps = ["AbstractFFTs", "Adapt", "CEnum", "CUDAapi", "CUDAdrv", "CUDAnative", "DataStructures", "GPUArrays", "Libdl", "LinearAlgebra", "MacroTools", "NNlib", "Pkg", "Printf", "Random", "Reexport", "Requires", "SparseArrays", "Statistics", "TimerOutputs"]
-git-tree-sha1 = "e8c55b38dcca955f5aed8ec4479cdc95810db1e1"
+git-tree-sha1 = "1582b74d2322df7dd94549d4ac9d095e0f20e884"
 uuid = "3a865a2d-5b23-5a0f-bc46-62713ec82fae"
-version = "2.0.1"
+version = "2.2.1"
 
 [[DataAPI]]
-git-tree-sha1 = "674b67f344687a88310213ddfa8a2b3c76cc4252"
+git-tree-sha1 = "176e23402d80e7743fc26c19c681bfb11246af32"
 uuid = "9a962f9c-6df0-11e9-0e5d-c546b8b5ee8a"
-version = "1.1.0"
+version = "1.3.0"
 
 [[DataStructures]]
 deps = ["InteractiveUtils", "OrderedCollections"]
-git-tree-sha1 = "73eb18320fe3ba58790c8b8f6f89420f0a622773"
+git-tree-sha1 = "af6d9c86e191c917c2276fbede1137e8ea20157f"
 uuid = "864edb3b-99cc-5e75-8d2d-829cb0a9cfe8"
-version = "0.17.11"
+version = "0.17.17"
 
 [[Dates]]
 deps = ["Printf"]
@@ -139,11 +139,16 @@ version = "1.0.1"
 deps = ["Random", "Serialization", "Sockets"]
 uuid = "8ba89e20-285c-5b6f-9357-94700520ee1b"
 
+[[ExprTools]]
+git-tree-sha1 = "6f0517056812fd6aa3af23d4b70d5325a2ae4e95"
+uuid = "e2ba6199-217a-4e67-a87a-7c52f15ade04"
+version = "0.1.1"
+
 [[FillArrays]]
 deps = ["LinearAlgebra", "Random", "SparseArrays"]
-git-tree-sha1 = "51cc2f9bc4eb9c6c0e81ec2f779d1085583cc956"
+git-tree-sha1 = "44f561e293987ffc84272cd3d2b14b0b93123d63"
 uuid = "1a297f60-69ca-5386-bcde-b61e274b549b"
-version = "0.8.7"
+version = "0.8.10"
 
 [[FixedPointNumbers]]
 git-tree-sha1 = "3ba9ea634d4c8b289d590403b4a06f8e227a6238"
@@ -162,17 +167,27 @@ git-tree-sha1 = "f40adc6422f548176bb4351ebd29e4abf773040a"
 uuid = "d9f16b24-f501-4c13-a1f2-28368ffc5196"
 version = "0.1.0"
 
+[[Future]]
+deps = ["Random"]
+uuid = "9fa8497b-333b-5362-9e8d-4d0656e87820"
+
 [[GPUArrays]]
 deps = ["AbstractFFTs", "Adapt", "LinearAlgebra", "Printf", "Random", "Serialization"]
-git-tree-sha1 = "d586762b08dcda13228df8967119b9cb6f22ade5"
+git-tree-sha1 = "d887693eb1bd5e1fd573262a978745481895ec7d"
 uuid = "0c68f7d7-f131-5f86-a1c3-88cf8149b2d7"
-version = "3.1.0"
+version = "3.4.1"
+
+[[GPUCompiler]]
+deps = ["Cthulhu", "DataStructures", "InteractiveUtils", "LLVM", "Libdl", "TimerOutputs"]
+git-tree-sha1 = "5275aa268ecd09640b32560e1eae90c78816e4d1"
+uuid = "61eb1bfa-7361-4325-ad38-22787b887f55"
+version = "0.2.0"
 
 [[IRTools]]
 deps = ["InteractiveUtils", "MacroTools", "Test"]
-git-tree-sha1 = "1a4355e4b5b50be2311ebb644f34f3306dbd0410"
+git-tree-sha1 = "90ee39f9beaaa186e4968417ea2b8ed5673c91c0"
 uuid = "7869d1d1-7146-5819-86e3-90919afe41df"
-version = "0.3.1"
+version = "0.3.3"
 
 [[InteractiveUtils]]
 deps = ["Markdown"]
@@ -180,15 +195,15 @@ uuid = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
 
 [[Juno]]
 deps = ["Base64", "Logging", "Media", "Profile"]
-git-tree-sha1 = "e1ba2a612645b3e07c773c3a208f215745081fe6"
+git-tree-sha1 = "a686b0cf235fa3e491b79b4783c2d2382292b436"
 uuid = "e5e0dc1b-0480-54bc-9374-aad01c23163d"
-version = "0.8.1"
+version = "0.8.2"
 
 [[LLVM]]
 deps = ["CEnum", "Libdl", "Printf", "Unicode"]
-git-tree-sha1 = "b6b86801ae2f2682e0a4889315dc76b68db2de71"
+git-tree-sha1 = "dd3f584c3dbefe39b2a8fbafa1a3b77e31e21255"
 uuid = "929cbde3-209d-540e-8aea-75f648917ca0"
-version = "1.3.4"
+version = "1.5.1"
 
 [[LibGit2]]
 deps = ["Printf"]
@@ -247,10 +262,9 @@ uuid = "efe28fd5-8261-553b-a9e1-b2916fc3738e"
 version = "0.5.3+3"
 
 [[OrderedCollections]]
-deps = ["Random", "Serialization", "Test"]
+git-tree-sha1 = "12ce190210d278e12644bcadf5b21cbdcf225cd3"
-git-tree-sha1 = "c4c13474d23c60d20a67b217f1d7f22a40edf8f1"
 uuid = "bac558e1-5e72-5ebc-8fee-abe8a469f55d"
-version = "1.1.0"
+version = "1.2.0"
 
 [[Pkg]]
 deps = ["Dates", "LibGit2", "Libdl", "Logging", "Markdown", "Printf", "REPL", "Random", "SHA", "UUIDs"]
@@ -305,15 +319,15 @@ uuid = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 
 [[SpecialFunctions]]
 deps = ["OpenSpecFun_jll"]
-git-tree-sha1 = "e19b98acb182567bcb7b75bb5d9eedf3a3b5ec6c"
+git-tree-sha1 = "d8d8b8a9f4119829410ecd706da4cc8594a1e020"
 uuid = "276daf66-3868-5448-9aa4-cd146d93841b"
-version = "0.10.0"
+version = "0.10.3"
 
 [[StaticArrays]]
 deps = ["LinearAlgebra", "Random", "Statistics"]
-git-tree-sha1 = "5a3bcb6233adabde68ebc97be66e95dcb787424c"
+git-tree-sha1 = "5c06c0aeb81bef54aed4b3f446847905eb6cbda0"
 uuid = "90137ffa-7385-5640-81b9-e52037218182"
-version = "0.12.1"
+version = "0.12.3"
 
 [[Statistics]]
 deps = ["LinearAlgebra", "SparseArrays"]
@@ -331,9 +345,9 @@ uuid = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 
 [[TimerOutputs]]
 deps = ["Printf"]
-git-tree-sha1 = "311765af81bbb48d7bad01fb016d9c328c6ede03"
+git-tree-sha1 = "f458ca23ff80e46a630922c555d838303e4b9603"
 uuid = "a759f4b9-e2f1-59dc-863e-4aeb61b1ea8f"
-version = "0.5.3"
+version = "0.5.6"
 
 [[TranscodingStreams]]
 deps = ["Random", "Test"]
@@ -350,21 +364,21 @@ uuid = "4ec0a83e-493e-50e2-b9ac-8f72acf5a8f5"
 
 [[ZipFile]]
 deps = ["Libdl", "Printf", "Zlib_jll"]
-git-tree-sha1 = "8748302cfdec02c4ae9c97b112cf10003f7f767f"
+git-tree-sha1 = "254975fef2fc526583bb9b7c9420fe66ffe09f2f"
 uuid = "a5390f91-8eb1-5f08-bee0-b1d1ffed6cea"
-version = "0.9.1"
+version = "0.9.2"
 
 [[Zlib_jll]]
 deps = ["Libdl", "Pkg"]
-git-tree-sha1 = "2f6c3e15e20e036ee0a0965879b31442b7ec50fa"
+git-tree-sha1 = "a2e0d558f6031002e380a90613b199e37a8565bf"
 uuid = "83775a58-1f1d-513f-b197-d71354ab007a"
-version = "1.2.11+9"
+version = "1.2.11+10"
 
 [[Zygote]]
-deps = ["AbstractFFTs", "ArrayLayouts", "DiffRules", "FillArrays", "ForwardDiff", "IRTools", "InteractiveUtils", "LinearAlgebra", "MacroTools", "NNlib", "NaNMath", "Random", "Requires", "SpecialFunctions", "Statistics", "ZygoteRules"]
+deps = ["AbstractFFTs", "ArrayLayouts", "DiffRules", "FillArrays", "ForwardDiff", "Future", "IRTools", "InteractiveUtils", "LinearAlgebra", "MacroTools", "NNlib", "NaNMath", "Random", "Requires", "SpecialFunctions", "Statistics", "ZygoteRules"]
-git-tree-sha1 = "1ccbfbe8930376e31752b812daa2532c723dc332"
+git-tree-sha1 = "707ceea58e2bd0ff3077ab13a92f8355181d3ee4"
 uuid = "e88e6eb3-aa80-5325-afca-941959d7151f"
-version = "0.4.13"
+version = "0.4.20"
 
 [[ZygoteRules]]
 deps = ["MacroTools"]

NEWS.md

@@ -1,5 +1,18 @@
+# v0.11
+* Change to `DataLoader`'s constructor [https://github.com/FluxML/Flux.jl/pull/1152]
+* Use `DataLoader` with `NamedTuple`s, so that tensors can be accessed by name [https://github.com/FluxML/Flux.jl/pull/1221].
+* Error if Dense layers weights and biases are not arrays [https://github.com/FluxML/Flux.jl/pull/1218].
+
 # v0.10.5
 * Add option for [same padding](https://github.com/FluxML/Flux.jl/pull/901) to conv and pooling layers by setting `pad=SamePad()`.
+* Added option to set `bias` to [Flux.Zeros](https://github.com/FluxML/Flux.jl/pull/873) to eliminating `bias` from being trained.
+* Added `GlobalMaxPool` and `GlobalMeanPool` [layers](https://github.com/FluxML/Flux.jl/pull/950) for performing global pooling operations.
+* Added `ClipValue` and `ClipNorm` in this [pr](https://github.com/FluxML/Flux.jl/pull/1133) to `Flux.Optimise` to provide a cleaner API for gradient clipping.
+* Added new kwarg-only [constructors](https://github.com/FluxML/Flux.jl/pull/873) for the various convolutional layers.
+* Documented the convolutional layer constructors accepting `weight` and `bias` keyword arguments to supply custom arrays for those fields.
+* Testing suite improvements now test for gradients of all layers along with GPU support.
+* Functors have now moved to [Functors.jl](https://github.com/FluxML/Flux.jl/pull/1174) to allow for their use outside of Flux.
+* Added [helper functions](https://github.com/FluxML/Flux.jl/pull/873) `Flux.convfilter` and `Flux.depthwiseconvfilter` to construct weight arrays for convolutions outside of layer constructors so as to not have to depend on the default layers for custom implementations.
 
 # v0.10.0
 * The default AD engine has switched from [Tracker to Zygote.jl](https://github.com/FluxML/Flux.jl/pull/669)

Project.toml

@@ -1,6 +1,6 @@
 name = "Flux"
 uuid = "587475ba-b771-5e3f-ad9e-33799f191a9c"
-version = "0.10.5"
+version = "0.11.0-DEV"
 
 [deps]
 AbstractTrees = "1520ce14-60c1-5f80-bbc7-55ef81b5835c"
@@ -11,6 +11,7 @@ CuArrays = "3a865a2d-5b23-5a0f-bc46-62713ec82fae"
 DelimitedFiles = "8bb1440f-4735-579b-a4ab-409b98df4dab"
 Functors = "d9f16b24-f501-4c13-a1f2-28368ffc5196"
 Juno = "e5e0dc1b-0480-54bc-9374-aad01c23163d"
+LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MacroTools = "1914dd2f-81c6-5fcd-8719-6d5c9610ff09"
 NNlib = "872c559c-99b0-510c-b3b7-b6c96a88d5cd"
 Pkg = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
@@ -26,10 +27,11 @@ Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [compat]
 AbstractTrees = "0.2, 0.3"
-Adapt = "1"
+Adapt = "1, 2.0"
 CodecZlib = "0.5, 0.6, 0.7"
 Colors = "0.8, 0.9, 0.10, 0.11, 0.12"
 CuArrays = "2"
+Functors = "0.1"
 Juno = "0.5, 0.6, 0.7, 0.8"
 MacroTools = "0.3, 0.4, 0.5"
 NNlib = "0.6"

docs/src/data/onehot.md

@@ -7,15 +7,15 @@ julia> using Flux: onehot, onecold
 
 julia> onehot(:b, [:a, :b, :c])
 3-element Flux.OneHotVector:
- false
+ 0
-  true
+ 1
- false
+ 0
 
 julia> onehot(:c, [:a, :b, :c])
 3-element Flux.OneHotVector:
- false
+ 0
- false
+ 0
-  true
+ 1
 ```
 
 The inverse is `onecold` (which can take a general probability distribution, as well as just booleans).

docs/src/models/advanced.md

@@ -19,7 +19,7 @@ Affine{Array{Float64,2},Array{Float64,1}}([0.66722 0.774872 0.249809; 0.843321 0
 julia> Flux.params(a) # default behavior
 Params([[0.66722 0.774872 0.249809; 0.843321 0.403843 0.429232; 0.683525 0.662455 0.065297], [0.42394, 0.0170927, 0.544955]])
 
-julia> Flux.trainable(a::Affine) = (a.W, a.b,)
+julia> Flux.trainable(a::Affine) = (a.W,)
 
 julia> Flux.params(a)
 Params([[0.66722 0.774872 0.249809; 0.843321 0.403843 0.429232; 0.683525 0.662455 0.065297]])

docs/src/models/layers.md

@@ -20,7 +20,11 @@ GlobalMeanPool
 DepthwiseConv
 ConvTranspose
 CrossCor
+SamePad
 flatten
+Flux.Zeros
+Flux.convfilter
+Flux.depthwiseconvfilter
 ```
 
 ## Recurrent Layers

docs/src/performance.md

@@ -39,7 +39,7 @@ E.g. the following will have run into the same problem as above:
     leaky_tanh(x) = 0.01*x + tanh(x)
 ```
 
-While one could change the activation function (e.g. to use `0.01f0x`), the idiomatic (and safe way)  to avoid type casts whenever inputs changes is to use `oftype`:
+While one could change the activation function (e.g. to use `0.01f0*x`), the idiomatic (and safe way)  to avoid type casts whenever inputs changes is to use `oftype`:
 ```
     leaky_tanh(x) = oftype(x/1, 0.01)*x + tanh(x)
 ```

docs/src/training/optimisers.md

@@ -140,3 +140,16 @@ ExpDecay
 InvDecay
 WeightDecay
 ```
+
+## Gradient Clipping
+
+Gradient clipping is useful for training recurrent neural networks, which have a tendency to suffer from the exploding gradient problem. An example usage is
+
+```julia
+opt = Optimiser(ClipValue(1e-3), ADAM(1e-3))
+```
+
+```@docs
+ClipValue
+ClipNorm
+```

docs/src/training/training.md

@@ -142,7 +142,7 @@ function my_custom_train!(loss, ps, data, opt)
   for d in data
     gs = gradient(ps) do
       training_loss = loss(d...)
-      # Insert what ever code you want here that needs Training loss, e.g. logging
+      # Insert whatever code you want here that needs Training loss, e.g. logging
       return training_loss
     end
     # insert what ever code you want here that needs gradient

src/Flux.jl

@@ -3,7 +3,8 @@ module Flux
 # Zero Flux Given
 
 using Base: tail
-using Zygote, MacroTools, Juno, Reexport, Statistics, Random
+using Statistics, Random, LinearAlgebra
+using Zygote, MacroTools, Juno, Reexport
 using MacroTools: @forward
 @reexport using NNlib
 using Zygote: Params, @adjoint, gradient, pullback, @nograd
@@ -20,7 +21,8 @@ using .Optimise
 using .Optimise: @epochs
 export Descent, ADAM, Momentum, Nesterov, RMSProp,
   ADAGrad, AdaMax, ADADelta, AMSGrad, NADAM,
-  ADAMW, RADAM, InvDecay, ExpDecay, WeightDecay
+  ADAMW, RADAM, InvDecay, ExpDecay, WeightDecay,
+  ClipValue, ClipNorm
 
 
 using CuArrays

src/data/Data.jl

@@ -51,4 +51,6 @@ export Iris
 include("housing.jl")
 export Housing
 
+@deprecate DataLoader(x...; kws...) DataLoader(x; kws...)
+
 end

src/data/dataloader.jl

@@ -1,7 +1,7 @@
 # Adapted from Knet's src/data.jl (author: Deniz Yuret)
 
-struct DataLoader
+struct DataLoader{D}
-    data
+    data::D
     batchsize::Int
     nobs::Int
     partial::Bool
@@ -11,21 +11,20 @@ struct DataLoader
 end
 
 """
-    DataLoader(data...; batchsize=1, shuffle=false, partial=true)
+    DataLoader(data; batchsize=1, shuffle=false, partial=true)
 
 An object that iterates over mini-batches of `data`, each mini-batch containing `batchsize` observations
 (except possibly the last one). 
 
-Takes as input one or more data tensors, e.g. X in unsupervised learning, X and Y in 
+Takes as input a single data tensor, or a tuple (or a named tuple) of tensors.
-supervised learning. The last dimension in each tensor is considered to be the observation
+The last dimension in each tensor is considered to be the observation dimension.
-dimension. 
 
 If `shuffle=true`, shuffles the observations each time iterations are re-started.
 If `partial=false`, drops the last mini-batch if it is smaller than the batchsize.
 
-The original data is preserved as a tuple in the `data` field of the DataLoader. 
+The original data is preserved in the `data` field of the DataLoader. 
 
-Example usage:
+Usage example:
 
     Xtrain = rand(10, 100)
     train_loader = DataLoader(Xtrain, batchsize=2) 
@@ -37,9 +36,16 @@ Example usage:
 
     train_loader.data   # original dataset
 
+    # similar, but yielding tuples
+    train_loader = DataLoader((Xtrain,), batchsize=2) 
+    for (x,) in train_loader
+        @assert size(x) == (10, 2)
+        ...
+    end
+
     Xtrain = rand(10, 100)
     Ytrain = rand(100)
-    train_loader = DataLoader(Xtrain, Ytrain, batchsize=2, shuffle=true) 
+    train_loader = DataLoader((Xtrain, Ytrain), batchsize=2, shuffle=true) 
     for epoch in 1:100
         for (x, y) in train_loader
             @assert size(x) == (10, 2)
@@ -51,26 +57,26 @@ Example usage:
     # train for 10 epochs
     using IterTools: ncycle 
     Flux.train!(loss, ps, ncycle(train_loader, 10), opt)
+
+    # can use NamedTuple to name tensors
+    train_loader = DataLoader((images=Xtrain, labels=Ytrain), batchsize=2, shuffle=true)
+    for datum in train_loader
+        @assert size(datum.images) == (10, 2)
+        @assert size(datum.labels) == (2,)
+    end
 """
-function DataLoader(data...; batchsize=1, shuffle=false, partial=true)
+function DataLoader(data; batchsize=1, shuffle=false, partial=true)
-    length(data) > 0 || throw(ArgumentError("Need at least one data input"))
     batchsize > 0 || throw(ArgumentError("Need positive batchsize"))
     
-    nx = size(data[1])[end]
+    n = _nobs(data) 
-    for i=2:length(data)
+    if n < batchsize
-        nx != size(data[i])[end] && throw(DimensionMismatch("All data should contain same number of observations"))
+        @warn "Number of observations less than batchsize, decreasing the batchsize to $n"
+        batchsize = n
     end
-    if nx < batchsize
+    imax = partial ? n : n - batchsize + 1
-        @warn "Number of data points less than batchsize, decreasing the batchsize to $nx"
+    DataLoader(data, batchsize, n, partial, imax, [1:n;], shuffle)
-        batchsize = nx
-    end
-    imax = partial ? nx : nx - batchsize + 1
-    ids = 1:min(nx, batchsize)
-    DataLoader(data, batchsize, nx, partial, imax, [1:nx;], shuffle)
 end
 
-getdata(x::AbstractArray, ids) = x[(Base.Colon() for _=1:ndims(x)-1)..., ids]
-
 @propagate_inbounds function Base.iterate(d::DataLoader, i=0)     # returns data in d.indices[i+1:i+batchsize]
     i >= d.imax && return nothing
     if d.shuffle && i == 0
@@ -78,11 +84,7 @@ getdata(x::AbstractArray, ids) = x[(Base.Colon() for _=1:ndims(x)-1)..., ids]
     end
     nexti = min(i + d.batchsize, d.nobs)
     ids = d.indices[i+1:nexti]
-    if length(d.data) == 1
+    batch = _getobs(d.data, ids)
-        batch = getdata(d.data[1], ids)
-    else
-        batch = ((getdata(x, ids) for x in d.data)...,)
-    end
     return (batch, nexti)
 end
 
@@ -90,3 +92,19 @@ function Base.length(d::DataLoader)
     n = d.nobs / d.batchsize
     d.partial ? ceil(Int,n) : floor(Int,n)
 end
+
+_nobs(data::AbstractArray) = size(data)[end]
+
+function _nobs(data::Union{Tuple, NamedTuple})
+    length(data) > 0 || throw(ArgumentError("Need at least one data input"))
+    n = _nobs(data[1])
+    if !all(x -> _nobs(x) == n, Base.tail(data))
+        throw(DimensionMismatch("All data should contain same number of observations"))
+    end
+    return n
+end
+
+_getobs(data::AbstractArray, i) = data[ntuple(i -> Colon(), Val(ndims(data) - 1))..., i]
+_getobs(data::Union{Tuple, NamedTuple}, i) = map(Base.Fix2(_getobs, i), data)
+
+Base.eltype(::DataLoader{D}) where D = D

src/functor.jl

@@ -24,7 +24,7 @@ testmode!(m, mode = true) = m
     trainmode!(m, mode = true)
 
 Set a layer of model's train mode (see below).
-Symmetric to [`testmode!`](@ref) (i.e. `trainmode!(m, mode) == testmode!(m, !mode)).
+Symmetric to [`testmode!`](@ref) (i.e. `trainmode!(m, mode) == testmode!(m, !mode)`).
 
 _Note_: if you manually set a model into train mode, you need to manually place
 it into test mode during testing phase.

src/layers/basic.jl

@@ -102,7 +102,7 @@ julia> d(rand(5))
   -0.16210233
    0.12311903```
 """
-struct Dense{F,S,T}
+struct Dense{F,S<:AbstractArray,T<:AbstractArray}
   W::S
   b::T
   σ::F

src/layers/conv.jl

@@ -132,7 +132,7 @@ end
 function (c::Conv)(x::AbstractArray)
   # TODO: breaks gpu broadcast :(
   # ndims(x) == ndims(c.weight)-1 && return squeezebatch(c(reshape(x, size(x)..., 1)))
-  σ, b = c.σ, reshape(c.bias, map(_->1, c.stride)..., :, 1)
+  σ, b = c.σ, reshape(c.bias, ntuple(_->1, length(c.stride))..., :, 1)
   cdims = DenseConvDims(x, c.weight; stride=c.stride, padding=c.pad, dilation=c.dilation)
   σ.(conv(x, c.weight, cdims) .+ b)
 end

src/layers/stateless.jl

@@ -46,9 +46,10 @@ given the prediction `ŷ` and true values `y`.
     Huber loss = |
                  |  δ * (|ŷ - y| - 0.5 * δ), otherwise
 """
+#TODO: remove dropgrad when Zygote can handle this function with CuArrays
 function huber_loss(ŷ, y;  δ=eltype(ŷ)(1))
    abs_error = abs.(ŷ .- y)
-   temp = abs_error .<  δ
+   temp = Zygote.dropgrad(abs_error .<  δ)
    x = eltype(ŷ)(0.5)
    hub_loss = sum(((abs_error.^2) .* temp) .* x .+ δ*(abs_error .- x*δ) .* (1 .- temp)) * 1 // length(y)
 end
@@ -132,7 +133,7 @@ CuArrays.@cufunc binarycrossentropy(ŷ, y; ϵ=eps(ŷ)) = -y*log(ŷ + ϵ) - (1
     logitbinarycrossentropy(ŷ, y)
 
 `logitbinarycrossentropy(ŷ, y)` is mathematically equivalent to
-[`Flux.binarycrossentropy(σ(log(ŷ)), y)`](@ref) but it is more numerically stable.
+[`Flux.binarycrossentropy(σ(ŷ), y)`](@ref) but it is more numerically stable.
 
 See also: [`Flux.crossentropy`](@ref), [`Flux.logitcrossentropy`](@ref), [`Flux.binarycrossentropy`](@ref)
 

src/onehot.jl

@@ -27,7 +27,8 @@ Base.getindex(xs::OneHotMatrix, ::Colon, ::Colon) = OneHotMatrix(xs.height, copy
 
 Base.getindex(xs::OneHotMatrix, i::Integer, ::Colon) = map(x -> x[i], xs.data)
 
-A::AbstractMatrix * B::OneHotMatrix = A[:, map(x->x.ix, B.data)]
+# remove workaround when https://github.com/JuliaGPU/CuArrays.jl/issues/676 is fixed
+A::AbstractMatrix * B::OneHotMatrix = A[:, cpu(map(x->x.ix, B.data))]
 
 Base.hcat(x::OneHotVector, xs::OneHotVector...) = OneHotMatrix(length(x), [x, xs...])
 
@@ -48,7 +49,7 @@ cudaconvert(x::OneHotMatrix{<:CuArray}) = OneHotMatrix(x.height, cudaconvert(x.d
 Create a `OneHotVector` with its `l`-th element `true` based on the
 possible set of `labels`.
 If `unk` is given, return `onehot(unk, labels)` if the input label `l` is not found
-in `labels`; otherwise it will error.
+in `labels`; otherwise, it will raise an error.
 
 # Examples
 ```jldoctest

src/optimise/Optimise.jl

@@ -1,9 +1,12 @@
 module Optimise
 
+using LinearAlgebra
+
 export train!, update!,
 	Descent, ADAM, Momentum, Nesterov, RMSProp,
 	ADAGrad, AdaMax, ADADelta, AMSGrad, NADAM, ADAMW,RADAM, 
-	InvDecay, ExpDecay, WeightDecay, stop, Optimiser
+	InvDecay, ExpDecay, WeightDecay, stop, Optimiser,
+	ClipValue, ClipNorm
 
 include("optimisers.jl")
 include("train.jl")

src/optimise/optimisers.jl

@@ -509,7 +509,7 @@ function apply!(o::ExpDecay, x, Δ)
   η, s, decay = o.eta, o.step, o.decay
   n = o.current[x] = get(o.current, x, 0) + 1
   if o.current[x]%s == 0 && count(x -> x%s == 0, values(o.current)) == 1
-    η = max(η * decay^(s / n), o.clip)
+    η = max(η * decay, o.clip)
     o.eta = η
   end
   @. Δ *= η
@@ -533,3 +533,31 @@ function apply!(o::WeightDecay, x, Δ)
   wd = o.wd
   @. Δ += wd * x
 end
+
+"""
+    ClipValue(thresh)
+
+Clip gradients when their absolute value exceeds `thresh`.
+"""
+mutable struct ClipValue{T}
+    thresh::T
+end
+
+apply!(o::ClipValue, x, Δ) = clamp!(Δ, -o.thresh, o.thresh)
+
+"""
+    ClipNorm(thresh)
+
+Clip gradients when their L2 norm exceeds `thresh`.
+"""
+mutable struct ClipNorm{T}
+    thresh::T
+end
+
+function apply!(o::ClipNorm, x, Δ)
+    Δnrm = norm(Δ)
+    if Δnrm > o.thresh
+        rmul!(Δ, o.thresh / Δnrm)
+    end
+    return Δ
+end

src/optimise/train.jl

@@ -68,8 +68,7 @@ and compute the gradient of `loss(d)`.
 A callback is given with the keyword argument `cb`. For example, this will print
 "training" every 10 seconds (using [`Flux.throttle`](@ref)):
 
-  train!(loss, params, data, opt,
+    train!(loss, params, data, opt, cb = throttle(() -> println("training"), 10))
-         cb = throttle(() -> println("training"), 10))
 
 The callback can call [`Flux.stop`](@ref) to interrupt the training loop.
 

src/utils.jl

@@ -246,6 +246,10 @@ function _restructure(m, xs)
   end
 end
 
+@adjoint function _restructure(m, xs)
+  _restructure(m, xs), dm -> (nothing,destructure(dm)[1])
+end
+
 """
     destructure(m)
 

test/cuda/cuda.jl

@@ -69,6 +69,7 @@ if CuArrays.has_cudnn()
   @info "Testing Flux/CUDNN"
   include("cudnn.jl")
   include("curnn.jl")
+  include("layers.jl")
 else
   @warn "CUDNN unavailable, not testing GPU DNN support"
 end

test/cuda/layers.jl

@@ -0,0 +1,98 @@
+# Test layers and data/model movements on and off the GPU
+# Add tests for layers and their gradients on the GPU
+# Most of the forward passes should be fine being applied
+# to bitstype objects, but this gives higher coverage for our use-cases
+# Check that getting the gradients does not throw
+
+# generic movement tests
+@testset "Basic GPU Movement" begin
+  @test gradient(x -> sum(gpu(x)), rand(3,3)) isa Tuple
+  @test gradient(x -> sum(cpu(x)), gpu(rand(3,3))) isa Tuple
+end
+
+# TODO: These layers get into scalar indexing
+# `AlphaDropout` throws a compilation error on GPUs,
+# whereas, the rest are scalar indexing issues.
+const BROKEN_LAYERS = [DepthwiseConv,
+		       AlphaDropout,
+                       InstanceNorm,
+                       GroupNorm]
+
+function gradtest(name::String, layers::Vector, xs = nothing, args...)
+  isnothing(xs) && error("Missing input to test the layers against.")
+  @testset "$name GPU grad tests" begin
+    for layer in layers
+      @testset "$layer GPU grad test" begin
+        l = gpu(layer(args...))
+        xs = gpu(xs)
+        if any(x -> isa(l, x), BROKEN_LAYERS)
+          ps = Flux.params(l)
+          @test_broken gradient(() -> sum(l(xs)), ps) isa Flux.Zygote.Grads
+        else
+          ps = Flux.params(l)
+          @test gradient(() -> sum(l(xs)), ps) isa Flux.Zygote.Grads
+          gs = gradient(() -> sum(l(xs)), ps)
+
+          # Handle pooling layers
+          if !isempty(ps)
+            @test gs[first(ps)] isa Flux.CuArrays.CuArray
+          end
+        end
+      end
+    end
+  end
+end
+
+# Repeats from Conv, CrossCor
+
+r = rand(Float32, 28, 28, 1, 1)
+conv_layers = [Conv, ConvTranspose, CrossCor, DepthwiseConv]
+gradtest("Conv", conv_layers, r, (2,2), 1=>3)
+
+pooling_layers = [MaxPool, MeanPool]
+gradtest("Pooling", pooling_layers, r, (2,2))
+
+dropout_layers = [Dropout, AlphaDropout]
+gradtest("Dropout", dropout_layers, r, 0.5f0)
+
+norm_layers = [LayerNorm, BatchNorm]
+gradtest("Normalising", norm_layers, rand(Float32, 28,28,3,1), 1)
+
+instancenorm = [InstanceNorm]
+gradtest("InstanceNorm", instancenorm, r, 1)
+
+groupnorm = [GroupNorm]
+gradtest("GroupNorm", groupnorm, rand(Float32, 28,28,3,1), 3, 1)
+
+const stateless_layers = [Flux.mse,
+                          Flux.crossentropy,
+                          Flux.logitcrossentropy,
+                          Flux.normalise]
+
+const stateless_layers_broadcasted = [Flux.binarycrossentropy,
+                                      Flux.logitbinarycrossentropy]
+
+function stateless_gradtest(f, args...)
+  @test gradient((args...) -> sum(f(args...)), args...)[1] isa CuArray
+end
+
+function stateless_gradtest_broadcasted(f, args...)
+  @test gradient((args...) -> sum(f.(args...)), args...)[1] isa CuArray
+end
+
+@testset "Stateless GPU grad tests" begin
+  x = gpu(rand(3,3))
+  y = gpu(rand(3,3))
+
+  for layer in stateless_layers
+    if layer == Flux.normalise
+      stateless_gradtest(layer, x)
+    else
+      stateless_gradtest(layer, x, y)
+    end
+  end
+
+  for layer in stateless_layers_broadcasted
+    stateless_gradtest_broadcasted(layer, x, y)
+  end
+end

test/data.jl

@@ -3,20 +3,34 @@
     Y = [1:5;]
 
     d = DataLoader(X, batchsize=2)
+    @inferred first(d)
     batches = collect(d)
+    @test eltype(batches) == eltype(d) == typeof(X)
     @test length(batches) == 3
     @test batches[1] == X[:,1:2]
     @test batches[2] == X[:,3:4]
     @test batches[3] == X[:,5:5]
 
     d = DataLoader(X, batchsize=2, partial=false)
+    @inferred first(d)
     batches = collect(d)
+    @test eltype(batches) == eltype(d) == typeof(X)
     @test length(batches) == 2
     @test batches[1] == X[:,1:2]
     @test batches[2] == X[:,3:4]
 
-    d = DataLoader(X, Y, batchsize=2)
+    d = DataLoader((X,), batchsize=2, partial=false)
+    @inferred first(d)
     batches = collect(d)
+    @test eltype(batches) == eltype(d) == Tuple{typeof(X)}
+    @test length(batches) == 2
+    @test batches[1] == (X[:,1:2],)
+    @test batches[2] == (X[:,3:4],)
+
+    d = DataLoader((X, Y), batchsize=2)
+    @inferred first(d)
+    batches = collect(d)
+    @test eltype(batches) == eltype(d) == Tuple{typeof(X), typeof(Y)}
     @test length(batches) == 3
     @test length(batches[1]) == 2
     @test length(batches[2]) == 2
@@ -28,6 +42,22 @@
     @test batches[3][1] == X[:,5:5]
     @test batches[3][2] == Y[5:5]
 
+    # test with NamedTuple
+    d = DataLoader((x=X, y=Y), batchsize=2)
+    @inferred first(d)
+    batches = collect(d)
+    @test eltype(batches) == eltype(d) == NamedTuple{(:x, :y), Tuple{typeof(X), typeof(Y)}}
+    @test length(batches) == 3
+    @test length(batches[1]) == 2
+    @test length(batches[2]) == 2
+    @test length(batches[3]) == 2
+    @test batches[1][1] == batches[1].x == X[:,1:2]
+    @test batches[1][2] == batches[1].y == Y[1:2]
+    @test batches[2][1] == batches[2].x == X[:,3:4]
+    @test batches[2][2] == batches[2].y == Y[3:4]
+    @test batches[3][1] == batches[3].x == X[:,5:5]
+    @test batches[3][2] == batches[3].y == Y[5:5]
+
     # test interaction with `train!`
     θ = ones(2)
     X = zeros(2, 10)
@@ -41,7 +71,7 @@
     X = ones(2, 10)
     Y = fill(2, 10)
     loss(x, y) = sum((y - x'*θ).^2)
-    d  = DataLoader(X, Y) 
+    d  = DataLoader((X, Y)) 
     Flux.train!(loss, [θ], ncycle(d, 10), Descent(0.1))
     @test norm(θ .- 1) < 1e-10
 end

test/layers/basic.jl

@@ -28,6 +28,14 @@ import Flux: activations
   end
 
   @testset "Dense" begin
+    @testset "constructors" begin
+      @test size(Dense(10, 100).W) == (100, 10)
+      @test Dense(rand(100,10), rand(10)).σ == identity
+
+      @test_throws MethodError Dense(10, 10.5)
+      @test_throws MethodError Dense(10, 10.5, tanh)
+    end
+
     @test  length(Dense(10, 5)(randn(10))) == 5
     @test_throws DimensionMismatch Dense(10, 5)(randn(1))
     @test_throws MethodError Dense(10, 5)(1) # avoid broadcasting
@@ -37,7 +45,6 @@ import Flux: activations
     @test Dense(10, 1, identity, initW = ones, initb = zeros)(ones(10,2)) == 10*ones(1, 2)
     @test Dense(10, 2, identity, initW = ones, initb = zeros)(ones(10,1)) == 10*ones(2, 1)
     @test Dense(10, 2, identity, initW = ones, initb = zeros)([ones(10,1) 2*ones(10,1)]) == [10 20; 10 20]
-
   end
 
   @testset "Diagonal" begin

test/optimise.jl

@@ -57,35 +57,57 @@ end
 end
 
 @testset "ExpDecay" begin
-    w = randn(10, 10)
+
-    o = ExpDecay(0.1, 0.1, 1000, 1e-4)
+  @testset "Sanity Check" begin
-    w1 = randn(10,10)
+    o = ExpDecay(0.2, 0.5, 1, 1e-3)
-    loss(x) = Flux.mse(w*x, w1*x)
+    p = [0.0]
-    flag = 1
+    steps = 1:8
-    decay_steps = []
+    eta_expected = @. max(o.eta * 0.5 ^ steps, o.clip)
-    for t = 1:10^5
+    eta_actual = [Optimise.apply!(o, p, [1.0])[1] for _ in steps]
-      prev_eta = o.eta
+    @test eta_actual == eta_expected
-      θ = Params([w1])
+  end
-      x = rand(10)
+
-      θ̄ = gradient(() -> loss(x), θ)
+  w = randn(10, 10)
-      prev_grad = collect(θ̄[w1])
+  o = ExpDecay(0.1, 0.1, 1000, 1e-4)
-      delta = Optimise.apply!(o, w1, θ̄[w1])
+  w1 = randn(10,10)
-      w1 .-= delta
+  loss(x) = Flux.mse(w*x, w1*x)
-      new_eta = o.eta
+  flag = 1
-      if new_eta != prev_eta
+  decay_steps = []
-        push!(decay_steps, t)
+  for t = 1:10^5
-      end
+    prev_eta = o.eta
-      array = fill(o.eta, size(prev_grad))
+    θ = Params([w1])
-      if array .* prev_grad != delta
+    x = rand(10)
-        flag = 0
+    θ̄ = gradient(() -> loss(x), θ)
-      end
+    prev_grad = collect(θ̄[w1])
+    delta = Optimise.apply!(o, w1, θ̄[w1])
+    w1 .-= delta
+    new_eta = o.eta
+    if new_eta != prev_eta
+      push!(decay_steps, t)
     end
-    @test flag == 1
+    array = fill(o.eta, size(prev_grad))
-    # Test to check if decay happens at decay steps. Eta reaches clip value eventually.
+    if array .* prev_grad != delta
-    ground_truth = []
+      flag = 0
-    for i in 1:11
-      push!(ground_truth, 1000*i)  # Expected decay steps for this example.
     end
-    @test decay_steps == ground_truth
+  end
-    @test o.eta == o.clip
+  @test flag == 1
+  # Test to check if decay happens at decay steps. Eta reaches clip value (1e-4) after 4000 steps (decay by 0.1 every 1000 steps starting at 0.1).
+  ground_truth = []
+  for i in 1:4
+    push!(ground_truth, 1000*i)  # Expected decay steps for this example.
+  end
+  @test decay_steps == ground_truth
+  @test o.eta == o.clip
+end
+
+@testset "Clipping" begin
+    w = randn(10, 10)
+    loss(x) = sum(w * x)
+    θ = Params([w])
+    x = 1000 * randn(10)
+    w̄ = gradient(() -> loss(x), θ)[w]
+    w̄_value = Optimise.apply!(ClipValue(1.0), w, copy(w̄))
+    @test all(w̄_value .<= 1)
+    w̄_norm = Optimise.apply!(ClipNorm(1.0), w, copy(w̄))
+    @test norm(w̄_norm) <= 1
 end

test/runtests.jl

@@ -2,49 +2,45 @@ using Flux
 using Flux.Data
 using Test 
 using Random, Statistics, LinearAlgebra
-using Documenter
 using IterTools: ncycle
 
 Random.seed!(0)
 
-@testset "Flux" begin
+@testset "Utils" begin
+  include("utils.jl")
+end
 
-  @testset "Utils" begin
+@testset "Onehot" begin
-    include("utils.jl")
+  include("onehot.jl")
-  end
+end
-
+
-  @testset "Onehot" begin
+@testset "Optimise" begin
-    include("onehot.jl")
+  include("optimise.jl")
-  end
+end
-
+
-  @testset "Optimise" begin
+@testset "Data" begin
-    include("optimise.jl")
+  include("data.jl")
-  end
+end
-
+
-  @testset "Data" begin
+@testset "Layers" begin
-    include("data.jl")
+  include("layers/basic.jl")
-  end
+  include("layers/normalisation.jl")
-
+  include("layers/stateless.jl")
-  @testset "Layers" begin
+  include("layers/conv.jl")
-    include("layers/basic.jl")
+end
-    include("layers/normalisation.jl")
+
-    include("layers/stateless.jl")
+@testset "CUDA" begin
-    include("layers/conv.jl")
+  if Flux.use_cuda[]
-  end
+    include("cuda/cuda.jl")
-
+  else
-  @testset "CUDA" begin
+    @warn "CUDA unavailable, not testing GPU support"
-    if Flux.use_cuda[]
-      include("cuda/cuda.jl")
-    else
-      @warn "CUDA unavailable, not testing GPU support"
-    end
   end
+end
 
+@static if VERSION >= v"1.4"
+  using Documenter
   @testset "Docs" begin
-    if VERSION >= v"1.4"
+    DocMeta.setdocmeta!(Flux, :DocTestSetup, :(using Flux); recursive=true)
-      DocMeta.setdocmeta!(Flux, :DocTestSetup, :(using Flux); recursive=true)
+    doctest(Flux)
-      doctest(Flux)
-    end
   end
-
+end
-end # testset Flux