diff --git a/.travis.yml b/.travis.yml index e02f470f..d2ab83c2 100644 --- a/.travis.yml +++ b/.travis.yml @@ -7,16 +7,16 @@ os: julia: - 1.3 + - 1 - nightly -matrix: - allow_failures: - - julia: nightly +notifications: + email: false jobs: include: - stage: "Documentation" - julia: 1.3 + julia: 1 os: linux script: - julia --project=docs/ -e 'using Pkg; Pkg.develop(PackageSpec(path=pwd())); @@ -24,6 +24,9 @@ jobs: - julia --project=docs/ docs/make.jl after_success: skip + allow_failures: + - julia: nightly + ## uncomment the following lines to override the default test script script: - julia --color=yes -e 'using Pkg; Pkg.activate(); Pkg.instantiate(); Pkg.test()' diff --git a/Manifest.toml b/Manifest.toml index 12986ccd..bc74cb6e 100644 --- a/Manifest.toml +++ b/Manifest.toml @@ -8,15 +8,21 @@ version = "0.5.0" [[AbstractTrees]] deps = ["Markdown"] -git-tree-sha1 = "8201f932428d25a2e2903300764515754847d87d" +git-tree-sha1 = "86d092c2599f1f7bb01668bf8eb3412f98d61e47" uuid = "1520ce14-60c1-5f80-bbc7-55ef81b5835c" -version = "0.3.0" +version = "0.3.2" [[Adapt]] deps = ["LinearAlgebra"] -git-tree-sha1 = "82dab828020b872fa9efd3abec1152b075bc7cbf" +git-tree-sha1 = "c88cfc7f9c1f9f8633cddf0b56e86302b70f64c5" uuid = "79e6a3ab-5dfb-504d-930d-738a2a938a0e" -version = "1.0.0" +version = "1.0.1" + +[[ArrayLayouts]] +deps = ["FillArrays", "LinearAlgebra"] +git-tree-sha1 = "41956a49a8a4fefa1bf6664bca4a3035aba4c3a0" +uuid = "4c555306-a7a7-4459-81d9-ec55ddd5c99a" +version = "0.2.3" [[Base64]] uuid = "2a0f44e3-6c83-55bd-87e4-b1978d98bd5f" @@ -34,39 +40,45 @@ version = "0.2.0" [[CUDAapi]] deps = ["Libdl", "Logging"] -git-tree-sha1 = "56a813440ac98a1aa64672ab460a1512552211a7" +git-tree-sha1 = "831b825d10104bd29e28f6da93312a976830717b" uuid = "3895d2a7-ec45-59b8-82bb-cfc6a382f9b3" -version = "2.1.0" +version = "4.0.0" [[CUDAdrv]] deps = ["CEnum", "CUDAapi", "Printf"] -git-tree-sha1 = "1fce616fa0806c67c133eb1d2f68f0f1a7504665" +git-tree-sha1 = "e650cbaee92b60433313157926b1e80d0c3a0e2e" uuid = "c5f51814-7f29-56b8-a69c-e4d8f6be1fde" -version = "5.0.1" +version = "6.2.2" [[CUDAnative]] -deps = ["Adapt", "CEnum", "CUDAapi", "CUDAdrv", "DataStructures", "InteractiveUtils", "LLVM", "Libdl", "Printf", "TimerOutputs"] -git-tree-sha1 = "6e11d5c2c91fc623952e94c4fb73f9c4db74795a" +deps = ["Adapt", "BinaryProvider", "CEnum", "CUDAapi", "CUDAdrv", "Cthulhu", "DataStructures", "InteractiveUtils", "LLVM", "Libdl", "MacroTools", "Pkg", "Printf", "TimerOutputs"] +git-tree-sha1 = "d1fc99635d0002c8a819b78cb1f441eb44310725" uuid = "be33ccc6-a3ff-5ff2-a52e-74243cff1e17" -version = "2.7.0" +version = "3.0.2" + +[[CodeTracking]] +deps = ["InteractiveUtils", "UUIDs"] +git-tree-sha1 = "0becdab7e6fbbcb7b88d8de5b72e5bb2f28239f3" +uuid = "da1fd8a2-8d9e-5ec2-8556-3022fb5608a2" +version = "0.5.8" [[CodecZlib]] -deps = ["BinaryProvider", "Libdl", "TranscodingStreams"] -git-tree-sha1 = "05916673a2627dd91b4969ff8ba6941bc85a960e" +deps = ["TranscodingStreams", "Zlib_jll"] +git-tree-sha1 = "ded953804d019afa9a3f98981d99b33e3db7b6da" uuid = "944b1d66-785c-5afd-91f1-9de20f533193" -version = "0.6.0" +version = "0.7.0" [[ColorTypes]] deps = ["FixedPointNumbers", "Random"] -git-tree-sha1 = "7b62b728a5f3dd6ee3b23910303ccf27e82fad5e" +git-tree-sha1 = "c4c1cca28748906265ed62c788d6fe6f0134d264" uuid = "3da002f7-5984-5a60-b8a6-cbb66c0b333f" -version = "0.8.1" +version = "0.10.0" [[Colors]] -deps = ["ColorTypes", "FixedPointNumbers", "InteractiveUtils", "Printf", "Reexport"] -git-tree-sha1 = "c9c1845d6bf22e34738bee65c357a69f416ed5d1" +deps = ["ColorTypes", "FixedPointNumbers", "InteractiveUtils", "Reexport"] +git-tree-sha1 = "2fdeb981ebcf52cd800ddb6a0aa5eac34153552d" uuid = "5ae59095-9a9b-59fe-a467-6f913c188581" -version = "0.9.6" +version = "0.12.0" [[CommonSubexpressions]] deps = ["Test"] @@ -74,11 +86,23 @@ git-tree-sha1 = "efdaf19ab11c7889334ca247ff4c9f7c322817b0" uuid = "bbf7d656-a473-5ed7-a52c-81e309532950" version = "0.2.0" +[[CompilerSupportLibraries_jll]] +deps = ["Libdl", "Pkg"] +git-tree-sha1 = "7c4f882c41faa72118841185afc58a2eb00ef612" +uuid = "e66e0078-7015-5450-92f7-15fbd957f2ae" +version = "0.3.3+0" + +[[Cthulhu]] +deps = ["CodeTracking", "InteractiveUtils", "REPL", "Unicode"] +git-tree-sha1 = "484790098c85c26f8e59051f8ff1a0745c034a7d" +uuid = "f68482b8-f384-11e8-15f7-abe071a5a75f" +version = "1.0.1" + [[CuArrays]] -deps = ["AbstractFFTs", "Adapt", "CEnum", "CUDAapi", "CUDAdrv", "CUDAnative", "DataStructures", "GPUArrays", "Libdl", "LinearAlgebra", "MacroTools", "NNlib", "Printf", "Random", "Requires", "SparseArrays", "TimerOutputs"] -git-tree-sha1 = "51fbe053dea29ed2513e02d38380007310cf4c4b" +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" uuid = "3a865a2d-5b23-5a0f-bc46-62713ec82fae" -version = "1.6.0" +version = "2.0.1" [[DataAPI]] git-tree-sha1 = "674b67f344687a88310213ddfa8a2b3c76cc4252" @@ -87,9 +111,9 @@ version = "1.1.0" [[DataStructures]] deps = ["InteractiveUtils", "OrderedCollections"] -git-tree-sha1 = "f784254f428fb8fd7ac15982e5862a38a44523d3" +git-tree-sha1 = "73eb18320fe3ba58790c8b8f6f89420f0a622773" uuid = "864edb3b-99cc-5e75-8d2d-829cb0a9cfe8" -version = "0.17.7" +version = "0.17.11" [[Dates]] deps = ["Printf"] @@ -107,78 +131,61 @@ version = "1.0.2" [[DiffRules]] deps = ["NaNMath", "Random", "SpecialFunctions"] -git-tree-sha1 = "10dca52cf6d4a62d82528262921daf63b99704a2" +git-tree-sha1 = "eb0c34204c8410888844ada5359ac8b96292cfd1" uuid = "b552c78f-8df3-52c6-915a-8e097449b14b" -version = "1.0.0" +version = "1.0.1" [[Distributed]] deps = ["Random", "Serialization", "Sockets"] uuid = "8ba89e20-285c-5b6f-9357-94700520ee1b" -[[FFTW]] -deps = ["AbstractFFTs", "FFTW_jll", "IntelOpenMP_jll", "Libdl", "LinearAlgebra", "MKL_jll", "Reexport"] -git-tree-sha1 = "109d82fa4b00429f9afcce873e9f746f11f018d3" -uuid = "7a1cc6ca-52ef-59f5-83cd-3a7055c09341" -version = "1.2.0" - -[[FFTW_jll]] -deps = ["Libdl", "Pkg"] -git-tree-sha1 = "05674f209a6e3387dd103a945b0113eeb64b1a58" -uuid = "f5851436-0d7a-5f13-b9de-f02708fd171a" -version = "3.3.9+3" - [[FillArrays]] deps = ["LinearAlgebra", "Random", "SparseArrays"] -git-tree-sha1 = "fec413d4fc547992eb62a5c544cedb6d7853c1f5" +git-tree-sha1 = "51cc2f9bc4eb9c6c0e81ec2f779d1085583cc956" uuid = "1a297f60-69ca-5386-bcde-b61e274b549b" -version = "0.8.4" +version = "0.8.7" [[FixedPointNumbers]] -git-tree-sha1 = "d14a6fa5890ea3a7e5dcab6811114f132fec2b4b" +git-tree-sha1 = "3ba9ea634d4c8b289d590403b4a06f8e227a6238" uuid = "53c48c17-4a7d-5ca2-90c5-79b7896eea93" -version = "0.6.1" +version = "0.8.0" [[ForwardDiff]] deps = ["CommonSubexpressions", "DiffResults", "DiffRules", "NaNMath", "Random", "SpecialFunctions", "StaticArrays"] -git-tree-sha1 = "840700059391d36e2498d89c2e82c08f261f2a2a" +git-tree-sha1 = "869540e4367122fbffaace383a5bdc34d6e5e5ac" uuid = "f6369f11-7733-5829-9624-2563aa707210" -version = "0.10.8" +version = "0.10.10" [[GPUArrays]] deps = ["AbstractFFTs", "Adapt", "LinearAlgebra", "Printf", "Random", "Serialization"] -git-tree-sha1 = "e756da6cee76a5f1436a05827fa8fdf3badc577f" +git-tree-sha1 = "d586762b08dcda13228df8967119b9cb6f22ade5" uuid = "0c68f7d7-f131-5f86-a1c3-88cf8149b2d7" -version = "2.0.1" +version = "3.1.0" [[IRTools]] deps = ["InteractiveUtils", "MacroTools", "Test"] -git-tree-sha1 = "72421971e60917b8cd7737f9577c4f0f87eab306" +git-tree-sha1 = "1a4355e4b5b50be2311ebb644f34f3306dbd0410" uuid = "7869d1d1-7146-5819-86e3-90919afe41df" -version = "0.3.0" - -[[IntelOpenMP_jll]] -deps = ["Libdl", "Pkg"] -git-tree-sha1 = "fb8e1c7a5594ba56f9011310790e03b5384998d6" -uuid = "1d5cc7b8-4909-519e-a0f8-d0f5ad9712d0" -version = "2018.0.3+0" +version = "0.3.1" [[InteractiveUtils]] deps = ["Markdown"] uuid = "b77e0a4c-d291-57a0-90e8-8db25a27a240" [[Juno]] -deps = ["Base64", "Logging", "Media", "Profile", "Test"] -git-tree-sha1 = "30d94657a422d09cb97b6f86f04f750fa9c50df8" +deps = ["Base64", "Logging", "Media", "Profile"] +git-tree-sha1 = "e1ba2a612645b3e07c773c3a208f215745081fe6" uuid = "e5e0dc1b-0480-54bc-9374-aad01c23163d" -version = "0.7.2" +version = "0.8.1" [[LLVM]] deps = ["CEnum", "Libdl", "Printf", "Unicode"] -git-tree-sha1 = "1d08d7e4250f452f6cb20e4574daaebfdbee0ff7" +git-tree-sha1 = "b6b86801ae2f2682e0a4889315dc76b68db2de71" uuid = "929cbde3-209d-540e-8aea-75f648917ca0" -version = "1.3.3" +version = "1.3.4" [[LibGit2]] +deps = ["Printf"] uuid = "76f85450-5226-5b5a-8eaa-529ad045b433" [[Libdl]] @@ -191,17 +198,11 @@ uuid = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e" [[Logging]] uuid = "56ddb016-857b-54e1-b83d-db4d58db5568" -[[MKL_jll]] -deps = ["Libdl", "Pkg"] -git-tree-sha1 = "61069ae718b8ab1e325bbfb4e5268902e7ea08e3" -uuid = "856f044c-d86e-5d09-b602-aeab76dc8ba7" -version = "2019.0.117+0" - [[MacroTools]] -deps = ["DataStructures", "Markdown", "Random"] -git-tree-sha1 = "e2fc7a55bb2224e203bbd8b59f72b91323233458" +deps = ["Markdown", "Random"] +git-tree-sha1 = "f7d2e3f654af75f01ec49be82c231c382214223a" uuid = "1914dd2f-81c6-5fcd-8719-6d5c9610ff09" -version = "0.5.3" +version = "0.5.5" [[Markdown]] deps = ["Base64"] @@ -224,9 +225,9 @@ uuid = "a63ad114-7e13-5084-954f-fe012c677804" [[NNlib]] deps = ["BinaryProvider", "Libdl", "LinearAlgebra", "Requires", "Statistics"] -git-tree-sha1 = "135c0de4794d5e214b06f1fb4787af4a72896e61" +git-tree-sha1 = "d9f196d911f55aeaff11b11f681b135980783824" uuid = "872c559c-99b0-510c-b3b7-b6c96a88d5cd" -version = "0.6.2" +version = "0.6.6" [[NaNMath]] git-tree-sha1 = "928b8ca9b2791081dc71a51c55347c27c618760f" @@ -234,10 +235,10 @@ uuid = "77ba4419-2d1f-58cd-9bb1-8ffee604a2e3" version = "0.3.3" [[OpenSpecFun_jll]] -deps = ["Libdl", "Pkg"] -git-tree-sha1 = "65f672edebf3f4e613ddf37db9dcbd7a407e5e90" +deps = ["CompilerSupportLibraries_jll", "Libdl", "Pkg"] +git-tree-sha1 = "d51c416559217d974a1113522d5919235ae67a87" uuid = "efe28fd5-8261-553b-a9e1-b2916fc3738e" -version = "0.5.3+1" +version = "0.5.3+3" [[OrderedCollections]] deps = ["Random", "Serialization", "Test"] @@ -273,9 +274,9 @@ version = "0.2.0" [[Requires]] deps = ["UUIDs"] -git-tree-sha1 = "999513b7dea8ac17359ed50ae8ea089e4464e35e" +git-tree-sha1 = "d37400976e98018ee840e0ca4f9d20baa231dc6b" uuid = "ae029012-a4dd-5104-9daa-d747884805df" -version = "1.0.0" +version = "1.0.1" [[SHA]] uuid = "ea8e919c-243c-51af-8825-aaa63cd721ce" @@ -298,9 +299,9 @@ uuid = "2f01184e-e22b-5df5-ae63-d93ebab69eaf" [[SpecialFunctions]] deps = ["OpenSpecFun_jll"] -git-tree-sha1 = "268052ee908b2c086cc0011f528694f02f3e2408" +git-tree-sha1 = "e19b98acb182567bcb7b75bb5d9eedf3a3b5ec6c" uuid = "276daf66-3868-5448-9aa4-cd146d93841b" -version = "0.9.0" +version = "0.10.0" [[StaticArrays]] deps = ["LinearAlgebra", "Random", "Statistics"] @@ -314,9 +315,9 @@ uuid = "10745b16-79ce-11e8-11f9-7d13ad32a3b2" [[StatsBase]] deps = ["DataAPI", "DataStructures", "LinearAlgebra", "Missings", "Printf", "Random", "SortingAlgorithms", "SparseArrays", "Statistics"] -git-tree-sha1 = "c53e809e63fe5cf5de13632090bc3520649c9950" +git-tree-sha1 = "a6102b1f364befdb05746f386b67c6b7e3262c45" uuid = "2913bbd2-ae8a-5f71-8c99-4fb6c76f3a91" -version = "0.32.0" +version = "0.33.0" [[Test]] deps = ["Distributed", "InteractiveUtils", "Logging", "Random"] @@ -343,21 +344,21 @@ uuid = "4ec0a83e-493e-50e2-b9ac-8f72acf5a8f5" [[ZipFile]] deps = ["Libdl", "Printf", "Zlib_jll"] -git-tree-sha1 = "5de8320a46812da1a8ca98b16a8a4546d44efa62" +git-tree-sha1 = "8748302cfdec02c4ae9c97b112cf10003f7f767f" uuid = "a5390f91-8eb1-5f08-bee0-b1d1ffed6cea" -version = "0.9.0" +version = "0.9.1" [[Zlib_jll]] deps = ["Libdl", "Pkg"] -git-tree-sha1 = "5618a43055eb09377edca21d19d0e99bce24a9c3" +git-tree-sha1 = "2f6c3e15e20e036ee0a0965879b31442b7ec50fa" uuid = "83775a58-1f1d-513f-b197-d71354ab007a" -version = "1.2.11+7" +version = "1.2.11+9" [[Zygote]] -deps = ["DiffRules", "FFTW", "FillArrays", "ForwardDiff", "IRTools", "InteractiveUtils", "LinearAlgebra", "MacroTools", "NNlib", "NaNMath", "Random", "Requires", "SpecialFunctions", "Statistics", "ZygoteRules"] -git-tree-sha1 = "74382bcc4c1e8075e14554da67d75565f8fb7827" +deps = ["AbstractFFTs", "ArrayLayouts", "DiffRules", "FillArrays", "ForwardDiff", "IRTools", "InteractiveUtils", "LinearAlgebra", "MacroTools", "NNlib", "NaNMath", "Random", "Requires", "SpecialFunctions", "Statistics", "ZygoteRules"] +git-tree-sha1 = "1ccbfbe8930376e31752b812daa2532c723dc332" uuid = "e88e6eb3-aa80-5325-afca-941959d7151f" -version = "0.4.5" +version = "0.4.13" [[ZygoteRules]] deps = ["MacroTools"] diff --git a/NEWS.md b/NEWS.md index 4023c7f2..460a9e5b 100644 --- a/NEWS.md +++ b/NEWS.md @@ -1,3 +1,6 @@ +# 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()`. + # v0.10.0 * The default AD engine has switched from [Tracker to Zygote.jl](https://github.com/FluxML/Flux.jl/pull/669) - The dependency on Tracker.jl has been removed. diff --git a/Project.toml b/Project.toml index 71282a10..1883d974 100644 --- a/Project.toml +++ b/Project.toml @@ -1,6 +1,6 @@ name = "Flux" uuid = "587475ba-b771-5e3f-ad9e-33799f191a9c" -version = "0.10.2" +version = "0.10.4" [deps] AbstractTrees = "1520ce14-60c1-5f80-bbc7-55ef81b5835c" @@ -26,21 +26,23 @@ Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f" [compat] AbstractTrees = "0.2, 0.3" Adapt = "1" -CodecZlib = "0.5, 0.6" -Colors = "0.8, 0.9, 0.10, 0.11" -CuArrays = "1.6" +CodecZlib = "0.5, 0.6, 0.7" +Colors = "0.8, 0.9, 0.10, 0.11, 0.12" +CuArrays = "2" Juno = "0.5, 0.6, 0.7, 0.8" MacroTools = "0.3, 0.4, 0.5" NNlib = "0.6" Reexport = "0.2" StatsBase = "0" ZipFile = "0.7, 0.8, 0.9" -Zygote = "0.4" -julia = "1" +Zygote = "0.4.13" +julia = "1.3" [extras] Documenter = "e30172f5-a6a5-5a46-863b-614d45cd2de4" +IterTools = "c8e1da08-722c-5040-9ed9-7db0dc04731e" +LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e" Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40" [targets] -test = ["Test", "Documenter"] +test = ["Test", "Documenter", "IterTools", "LinearAlgebra"] diff --git a/docs/Manifest.toml b/docs/Manifest.toml deleted file mode 100644 index bf9d220a..00000000 --- a/docs/Manifest.toml +++ /dev/null @@ -1,89 +0,0 @@ -# This file is machine-generated - editing it directly is not advised - -[[Base64]] -uuid = "2a0f44e3-6c83-55bd-87e4-b1978d98bd5f" - -[[Dates]] -deps = ["Printf"] -uuid = "ade2ca70-3891-5945-98fb-dc099432e06a" - -[[Distributed]] -deps = ["Random", "Serialization", "Sockets"] -uuid = "8ba89e20-285c-5b6f-9357-94700520ee1b" - -[[DocStringExtensions]] -deps = ["LibGit2", "Markdown", "Pkg", "Test"] -git-tree-sha1 = "0513f1a8991e9d83255e0140aace0d0fc4486600" -uuid = "ffbed154-4ef7-542d-bbb7-c09d3a79fcae" -version = "0.8.0" - -[[Documenter]] -deps = ["Base64", "DocStringExtensions", "InteractiveUtils", "JSON", "LibGit2", "Logging", "Markdown", "REPL", "Test", "Unicode"] -git-tree-sha1 = "c61d6eedbc3c4323c08b64af12d29c8ee0fcbb5f" -uuid = "e30172f5-a6a5-5a46-863b-614d45cd2de4" -version = "0.23.2" - -[[InteractiveUtils]] -deps = ["Markdown"] -uuid = "b77e0a4c-d291-57a0-90e8-8db25a27a240" - -[[JSON]] -deps = ["Dates", "Mmap", "Parsers", "Unicode"] -git-tree-sha1 = "b34d7cef7b337321e97d22242c3c2b91f476748e" -uuid = "682c06a0-de6a-54ab-a142-c8b1cf79cde6" -version = "0.21.0" - -[[LibGit2]] -uuid = "76f85450-5226-5b5a-8eaa-529ad045b433" - -[[Logging]] -uuid = "56ddb016-857b-54e1-b83d-db4d58db5568" - -[[Markdown]] -deps = ["Base64"] -uuid = "d6f4376e-aef5-505a-96c1-9c027394607a" - -[[Mmap]] -uuid = "a63ad114-7e13-5084-954f-fe012c677804" - -[[Parsers]] -deps = ["Dates", "Test"] -git-tree-sha1 = "db2b35dedab3c0e46dc15996d170af07a5ab91c9" -uuid = "69de0a69-1ddd-5017-9359-2bf0b02dc9f0" -version = "0.3.6" - -[[Pkg]] -deps = ["Dates", "LibGit2", "Markdown", "Printf", "REPL", "Random", "SHA", "UUIDs"] -uuid = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f" - -[[Printf]] -deps = ["Unicode"] -uuid = "de0858da-6303-5e67-8744-51eddeeeb8d7" - -[[REPL]] -deps = ["InteractiveUtils", "Markdown", "Sockets"] -uuid = "3fa0cd96-eef1-5676-8a61-b3b8758bbffb" - -[[Random]] -deps = ["Serialization"] -uuid = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c" - -[[SHA]] -uuid = "ea8e919c-243c-51af-8825-aaa63cd721ce" - -[[Serialization]] -uuid = "9e88b42a-f829-5b0c-bbe9-9e923198166b" - -[[Sockets]] -uuid = "6462fe0b-24de-5631-8697-dd941f90decc" - -[[Test]] -deps = ["Distributed", "InteractiveUtils", "Logging", "Random"] -uuid = "8dfed614-e22c-5e08-85e1-65c5234f0b40" - -[[UUIDs]] -deps = ["Random", "SHA"] -uuid = "cf7118a7-6976-5b1a-9a39-7adc72f591a4" - -[[Unicode]] -uuid = "4ec0a83e-493e-50e2-b9ac-8f72acf5a8f5" diff --git a/docs/Project.toml b/docs/Project.toml index dfa65cd1..670a65be 100644 --- a/docs/Project.toml +++ b/docs/Project.toml @@ -1,2 +1,6 @@ [deps] Documenter = "e30172f5-a6a5-5a46-863b-614d45cd2de4" +NNlib = "872c559c-99b0-510c-b3b7-b6c96a88d5cd" + +[compat] +Documenter = "0.24" diff --git a/docs/make.jl b/docs/make.jl index b950e959..2f24a022 100644 --- a/docs/make.jl +++ b/docs/make.jl @@ -1,29 +1,36 @@ -using Pkg; -Pkg.activate(joinpath(@__DIR__, "..")); Pkg.instantiate() -Pkg.activate(); Pkg.instantiate() - -pushfirst!(LOAD_PATH, joinpath(@__DIR__, "..")) - using Documenter, Flux, NNlib +DocMeta.setdocmeta!(Flux, :DocTestSetup, :(using Flux); recursive=true) makedocs(modules=[Flux, NNlib], + doctest = VERSION >= v"1.4", sitename = "Flux", pages = ["Home" => "index.md", "Building Models" => ["Basics" => "models/basics.md", "Recurrence" => "models/recurrence.md", "Regularisation" => "models/regularisation.md", - "Model Reference" => "models/layers.md"], + "Model Reference" => "models/layers.md", + "Advanced Model Building" => "models/advanced.md", + "NNlib" => "models/nnlib.md"], + "Handling Data" => + ["One-Hot Encoding" => "data/onehot.md", + "DataLoader" => "data/dataloader.md"], "Training Models" => ["Optimisers" => "training/optimisers.md", "Training" => "training/training.md"], - "One-Hot Encoding" => "data/onehot.md", "GPU Support" => "gpu.md", "Saving & Loading" => "saving.md", + "The Julia Ecosystem" => "ecosystem.md", + "Utility Functions" => "utilities.md", "Performance Tips" => "performance.md", + "Datasets" => "datasets.md", "Community" => "community.md"], - format = Documenter.HTML(assets = ["assets/flux.css"], - analytics = "UA-36890222-9", - prettyurls = haskey(ENV, "CI"))) + format = Documenter.HTML( + analytics = "UA-36890222-9", + assets = ["assets/flux.css"], + prettyurls = get(ENV, "CI", nothing) == "true"), + ) -deploydocs(repo = "github.com/FluxML/Flux.jl.git") +deploydocs(repo = "github.com/FluxML/Flux.jl.git", + target = "build", + push_preview = true) diff --git a/docs/src/data/dataloader.md b/docs/src/data/dataloader.md new file mode 100644 index 00000000..f6edc709 --- /dev/null +++ b/docs/src/data/dataloader.md @@ -0,0 +1,6 @@ +# DataLoader +Flux provides the `DataLoader` type in the `Flux.Data` module to handle iteration over mini-batches of data. + +```@docs +Flux.Data.DataLoader +``` diff --git a/docs/src/data/onehot.md b/docs/src/data/onehot.md index 0bc3531b..23d6f196 100644 --- a/docs/src/data/onehot.md +++ b/docs/src/data/onehot.md @@ -31,6 +31,11 @@ julia> onecold([0.3, 0.2, 0.5], [:a, :b, :c]) :c ``` +```@docs +Flux.onehot +Flux.onecold +``` + ## Batches `onehotbatch` creates a batch (matrix) of one-hot vectors, and `onecold` treats matrices as batches. @@ -52,3 +57,7 @@ julia> onecold(ans, [:a, :b, :c]) ``` Note that these operations returned `OneHotVector` and `OneHotMatrix` rather than `Array`s. `OneHotVector`s behave like normal vectors but avoid any unnecessary cost compared to using an integer index directly. For example, multiplying a matrix with a one-hot vector simply slices out the relevant row of the matrix under the hood. + +```@docs +Flux.onehotbatch +``` diff --git a/docs/src/datasets.md b/docs/src/datasets.md new file mode 100644 index 00000000..45e29a75 --- /dev/null +++ b/docs/src/datasets.md @@ -0,0 +1,20 @@ +# Datasets + +Flux includes several standard machine learning datasets. + +```@docs +Flux.Data.Iris.features() +Flux.Data.Iris.labels() +Flux.Data.MNIST.images() +Flux.Data.MNIST.labels() +Flux.Data.FashionMNIST.images() +Flux.Data.FashionMNIST.labels() +Flux.Data.CMUDict.phones() +Flux.Data.CMUDict.symbols() +Flux.Data.CMUDict.rawdict() +Flux.Data.CMUDict.cmudict() +Flux.Data.Sentiment.train() +Flux.Data.Sentiment.test() +Flux.Data.Sentiment.dev() +``` + diff --git a/docs/src/ecosystem.md b/docs/src/ecosystem.md new file mode 100644 index 00000000..0672ffe6 --- /dev/null +++ b/docs/src/ecosystem.md @@ -0,0 +1,21 @@ +# The Julia Ecosystem + +One of the main strengths of Julia lies in an ecosystem of packages +globally providing a rich and consistent user experience. + +This is a non-exhaustive list of Julia packages, nicely complementing `Flux` in typical +machine learning and deep learning workflows: + +- [ArgParse.jl](https://github.com/carlobaldassi/ArgParse.jl): package for parsing command-line arguments to Julia programs. +- [Augmentor.jl](https://github.com/Evizero/Augmentor.jl): a fast image augmentation library in Julia for machine learning. +- [BSON.jl](https://github.com/JuliaIO/BSON.jl): package for working with the Binary JSON serialisation format +- [DataFrames.jl](https://github.com/joshday/OnlineStats.jl): in-memory tabular data in Julia +- [DrWatson.jl](https://github.com/JuliaDynamics/DrWatson.jl): a scientific project assistant software +- [MLDatasets.jl](https://github.com/JuliaML/MLDatasets.jl): utility package for accessing common machine learning datasets +- [OnlineStats.jl](https://github.com/joshday/OnlineStats.jl): single-pass algorithms for statistics +- [Parameters.jl](https://github.com/mauro3/Parameters.jl): types with default field values, keyword constructors and (un-)pack macros +- [ProgressMeters.jl](https://github.com/timholy/ProgressMeter.jl): progress meters for long-running computations +- [TensorBoardLogger.jl](https://github.com/PhilipVinc/TensorBoardLogger.jl): easy peasy logging to [tensorboard](https://www.tensorflow.org/tensorboard) in Julia + + +This tight integration among Julia pakages is shown in some of the examples in the [model-zoo](https://github.com/FluxML/model-zoo) repository. \ No newline at end of file diff --git a/docs/src/gpu.md b/docs/src/gpu.md index bb13fdd1..19d0c8c6 100644 --- a/docs/src/gpu.md +++ b/docs/src/gpu.md @@ -30,7 +30,7 @@ If you define a structured model, like a `Dense` layer or `Chain`, you just need ```julia d = Dense(10, 5, σ) d = fmap(cu, d) -d.W # Tracked CuArray +d.W # CuArray d(cu(rand(10))) # CuArray output m = Chain(Dense(10, 5, σ), Dense(5, 2), softmax) @@ -53,7 +53,7 @@ julia> x = rand(10) |> gpu 0.511655 julia> m(x) -Tracked 5-element CuArray{Float32,1}: +5-element CuArray{Float32,1}: -0.30535 ⋮ -0.618002 diff --git a/docs/src/models/advanced.md b/docs/src/models/advanced.md new file mode 100644 index 00000000..cbdb226c --- /dev/null +++ b/docs/src/models/advanced.md @@ -0,0 +1,73 @@ +# Advanced Model Building and Customisation + +Here we will try and describe usage of some more advanced features that Flux provides to give more control over model building. + +## Customising Parameter Collection for a Model + +Taking reference from our example `Affine` layer from the [basics](basics.md#Building-Layers-1). + +By default all the fields in the `Affine` type are collected as its parameters, however, in some cases it may be desired to hold other metadata in our "layers" that may not be needed for training, and are hence supposed to be ignored while the parameters are collected. With Flux, it is possible to mark the fields of our layers that are trainable in two ways. + +The first way of achieving this is through overloading the `trainable` function. + +```julia-repl +julia> @functor Affine + +julia> a = Affine(rand(3,3), rand(3)) +Affine{Array{Float64,2},Array{Float64,1}}([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.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.params(a) +Params([[0.66722 0.774872 0.249809; 0.843321 0.403843 0.429232; 0.683525 0.662455 0.065297]]) +``` + +Only the fields returned by `trainable` will be collected as trainable parameters of the layer when calling `Flux.params`. + +Another way of achieving this is through the `@functor` macro directly. Here, we can mark the fields we are interested in by grouping them in the second argument: + +```julia +Flux.@functor Affine (W,) +``` + +However, doing this requires the `struct` to have a corresponding constructor that accepts those parameters. + +## Freezing Layer Parameters + +When it is desired to not include all the model parameters (for e.g. transfer learning), we can simply not pass in those layers into our call to `params`. + +Consider a simple multi-layer perceptron model where we want to avoid optimising the first two `Dense` layers. We can obtain +this using the slicing features `Chain` provides: + +```julia +m = Chain( + Dense(784, 64, relu), + Dense(64, 64, relu), + Dense(32, 10) + ) + +ps = Flux.params(m[3:end]) +``` + +The `Zygote.Params` object `ps` now holds a reference to only the parameters of the layers passed to it. + +During training, the gradients will only be computed for (and applied to) the last `Dense` layer, therefore only that would have its parameters changed. + +`Flux.params` also takes multiple inputs to make it easy to collect parameters from heterogenous models with a single call. A simple demonstration would be if we wanted to omit optimising the second `Dense` layer in the previous example. It would look something like this: + +```julia +Flux.params(m[1], m[3:end]) +``` + +Sometimes, a more fine-tuned control is needed. +We can freeze a specific parameter of a specific layer which already entered a `Params` object `ps`, +by simply deleting it from `ps`: + +```julia +ps = params(m) +delete!(ps, m[2].b) +``` + diff --git a/docs/src/models/basics.md b/docs/src/models/basics.md index 6e8d0b76..06901d99 100644 --- a/docs/src/models/basics.md +++ b/docs/src/models/basics.md @@ -69,8 +69,8 @@ b = rand(2) predict(x) = W*x .+ b function loss(x, y) - ŷ = predict(x) - sum((y .- ŷ).^2) + ŷ = predict(x) + sum((y .- ŷ).^2) end x, y = rand(5), rand(2) # Dummy data @@ -220,6 +220,8 @@ Flux.@functor Affine This enables a useful extra set of functionality for our `Affine` layer, such as [collecting its parameters](../training/optimisers.md) or [moving it to the GPU](../gpu.md). +For some more helpful tricks, including parameter freezing, please checkout the [advanced usage guide](advanced.md). + ## Utility functions Flux provides some utility functions to help you generate models in an automated fashion. @@ -238,5 +240,5 @@ Currently limited to the following layers: - `MeanPool` ```@docs -outdims +Flux.outdims ``` diff --git a/docs/src/models/layers.md b/docs/src/models/layers.md index 5f2ab3ce..54ce5791 100644 --- a/docs/src/models/layers.md +++ b/docs/src/models/layers.md @@ -14,10 +14,13 @@ These layers are used to build convolutional neural networks (CNNs). ```@docs Conv MaxPool +GlobalMaxPool MeanPool +GlobalMeanPool DepthwiseConv ConvTranspose CrossCor +flatten ``` ## Recurrent Layers @@ -29,6 +32,7 @@ RNN LSTM GRU Flux.Recur +Flux.reset! ``` ## Other General Purpose Layers @@ -40,40 +44,45 @@ Maxout SkipConnection ``` -## Activation Functions - -Non-linearities that go between layers of your model. Most of these functions are defined in [NNlib](https://github.com/FluxML/NNlib.jl) but are available by default in Flux. - -Note that, unless otherwise stated, activation functions operate on scalars. To apply them to an array you can call `σ.(xs)`, `relu.(xs)` and so on. - -```@docs -σ -relu -leakyrelu -elu -swish -``` ## Normalisation & Regularisation These layers don't affect the structure of the network but may improve training times or reduce overfitting. ```@docs +Flux.normalise BatchNorm +Flux.dropout Dropout AlphaDropout LayerNorm +InstanceNorm GroupNorm ``` +### Testmode + +Many normalisation layers behave differently under training and inference (testing). By default, Flux will automatically determine when a layer evaluation is part of training or inference. Still, depending on your use case, it may be helpful to manually specify when these layers should be treated as being trained or not. For this, Flux provides `Flux.testmode!`. When called on a model (e.g. a layer or chain of layers), this function will place the model into the mode specified. + +```@docs +Flux.testmode! +trainmode! +``` + ## Cost Functions ```@docs -mse -crossentropy -logitcrossentropy -binarycrossentropy -logitbinarycrossentropy -kldivergence -poisson -hinge +Flux.mae +Flux.mse +Flux.msle +Flux.huber_loss +Flux.crossentropy +Flux.logitcrossentropy +Flux.binarycrossentropy +Flux.logitbinarycrossentropy +Flux.kldivergence +Flux.poisson +Flux.hinge +Flux.squared_hinge +Flux.dice_coeff_loss +Flux.tversky_loss ``` diff --git a/docs/src/models/nnlib.md b/docs/src/models/nnlib.md new file mode 100644 index 00000000..7ede2682 --- /dev/null +++ b/docs/src/models/nnlib.md @@ -0,0 +1,61 @@ +# NNlib + +Flux re-exports all of the functions exported by the [NNlib](https://github.com/FluxML/NNlib.jl) package. + +## Activation Functions + +Non-linearities that go between layers of your model. Note that, unless otherwise stated, activation functions operate on scalars. To apply them to an array you can call `σ.(xs)`, `relu.(xs)` and so on. + +```@docs +NNlib.celu +NNlib.elu +NNlib.gelu +NNlib.hardsigmoid +NNlib.hardtanh +NNlib.leakyrelu +NNlib.lisht +NNlib.logcosh +NNlib.logsigmoid +NNlib.mish +NNlib.relu +NNlib.relu6 +NNlib.rrelu +NNlib.selu +NNlib.sigmoid +NNlib.softplus +NNlib.softshrink +NNlib.softsign +NNlib.swish +NNlib.tanhshrink +NNlib.trelu +``` + +## Softmax + +```@docs +NNlib.softmax +NNlib.logsoftmax +``` + +## Pooling + +```@docs +NNlib.maxpool +NNlib.meanpool +``` + +## Convolution + +```@docs +NNlib.conv +NNlib.depthwiseconv +``` + +## Batched Operations + +```@docs +NNlib.batched_mul +NNlib.batched_mul! +NNlib.batched_adjoint +NNlib.batched_transpose +``` diff --git a/docs/src/models/regularisation.md b/docs/src/models/regularisation.md index e1d88d77..535dd096 100644 --- a/docs/src/models/regularisation.md +++ b/docs/src/models/regularisation.md @@ -31,7 +31,7 @@ julia> params(m) param([0.0, 0.0, 0.0, 0.0, 0.0]) julia> sum(norm, params(m)) -26.01749952921026 (tracked) +26.01749952921026 ``` Here's a larger example with a multi-layer perceptron. @@ -52,7 +52,7 @@ One can also easily add per-layer regularisation via the `activations` function: ```julia julia> using Flux: activations -julia> c = Chain(Dense(10,5,σ),Dense(5,2),softmax) +julia> c = Chain(Dense(10, 5, σ), Dense(5, 2), softmax) Chain(Dense(10, 5, σ), Dense(5, 2), softmax) julia> activations(c, rand(10)) @@ -64,3 +64,7 @@ julia> activations(c, rand(10)) julia> sum(norm, ans) 2.1166067f0 ``` + +```@docs +Flux.activations +``` diff --git a/docs/src/performance.md b/docs/src/performance.md index 06a4f690..4601e90c 100644 --- a/docs/src/performance.md +++ b/docs/src/performance.md @@ -4,7 +4,7 @@ All the usual [Julia performance tips apply](https://docs.julialang.org/en/v1/ma As always [profiling your code](https://docs.julialang.org/en/v1/manual/profile/#Profiling-1) is generally a useful way of finding bottlenecks. Below follow some Flux specific tips/reminders. -## Don't use more precision than you need. +## Don't use more precision than you need Flux works great with all kinds of number types. But often you do not need to be working with say `Float64` (let alone `BigFloat`). @@ -14,7 +14,8 @@ Which means allocations occur much faster. And you use less memory. -## Make sure your activation and loss functions preserve the type of their inputs +## Preserve inputs' types + Not only should your activation and loss functions be [type-stable](https://docs.julialang.org/en/v1/manual/performance-tips/#Write-%22type-stable%22-functions-1), they should also preserve the type of their inputs. @@ -29,31 +30,29 @@ because it results in having to use slow mixed type multiplication in the dense Similar situations can occur in the loss function during backpropagation. Which means if you change your data say from `Float64` to `Float32` (which should give a speedup: see above), -you will see a large slow-down +you will see a large slow-down. This can occur sneakily, because you can cause type-promotion by interacting with a numeric literals. E.g. the following will have run into the same problem as above: ``` - leaky_tanh(x) = 0.01x + tanh(x) + leaky_tanh(x) = 0.01*x + tanh(x) ``` -While one could change your activation function (e.g. to use `0.01f0x`) to avoid this when ever your inputs change, -the idiomatic (and safe way) is to use `oftype`. - +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`: ``` - leaky_tanh(x) = oftype(x/1, 0.01)x + tanh(x) + leaky_tanh(x) = oftype(x/1, 0.01)*x + tanh(x) ``` -## Evaluate batches as Matrices of features, rather than sequences of Vector features +## Evaluate batches as Matrices of features While it can sometimes be tempting to process your observations (feature vectors) one at a time e.g. ```julia function loss_total(xs::AbstractVector{<:Vector}, ys::AbstractVector{<:Vector}) sum(zip(xs, ys)) do (x, y_target) - y_pred = model(x) # evaluate the model + y_pred = model(x) # evaluate the model return loss(y_pred, y_target) end end diff --git a/docs/src/training/optimisers.md b/docs/src/training/optimisers.md index 5e8b95de..5ed083ee 100644 --- a/docs/src/training/optimisers.md +++ b/docs/src/training/optimisers.md @@ -21,7 +21,7 @@ grads = gradient(() -> loss(x, y), θ) We want to update each parameter, using the gradient, in order to improve (reduce) the loss. Here's one way to do that: ```julia -using Flux: update! +using Flux.Optimise: update! η = 0.1 # Learning Rate for p in (W, b) @@ -46,11 +46,13 @@ An optimiser `update!` accepts a parameter and a gradient, and updates the param All optimisers return an object that, when passed to `train!`, will update the parameters passed to it. ```@docs +Flux.Optimise.update! Descent Momentum Nesterov RMSProp ADAM +RADAM AdaMax ADAGrad ADADelta @@ -61,7 +63,7 @@ ADAMW ## Optimiser Interface -Flux's optimsers are built around a `struct` that holds all the optimiser parameters along with a definition of how to apply the update rule associated with it. We do this via the `apply!` function which takes the optimiser as the first argument followed by the parameter and its corresponding gradient. +Flux's optimisers are built around a `struct` that holds all the optimiser parameters along with a definition of how to apply the update rule associated with it. We do this via the `apply!` function which takes the optimiser as the first argument followed by the parameter and its corresponding gradient. In this manner Flux also allows one to create custom optimisers to be used seamlessly. Let's work this with a simple example. @@ -99,15 +101,15 @@ Flux internally calls on this function via the `update!` function. It shares the ## Composing Optimisers -Flux defines a special kind of optimiser called simply as `Optimiser` which takes in a arbitrary optimisers as input. Its behaviour is similar to the usual optimisers, but differs in that it acts by calling the optimisers listed in it sequentially. Each optimiser produces a modified gradient +Flux defines a special kind of optimiser simply called `Optimiser` which takes in arbitrary optimisers as input. Its behaviour is similar to the usual optimisers, but differs in that it acts by calling the optimisers listed in it sequentially. Each optimiser produces a modified gradient that will be fed into the next, and the resultant update will be applied to the parameter as usual. A classic use case is where adding decays is desirable. Flux defines some basic decays including `ExpDecay`, `InvDecay` etc. ```julia opt = Optimiser(ExpDecay(0.001, 0.1, 1000, 1e-4), Descent()) ``` -Here we apply exponential decay to the `Descent` optimser. The defaults of `ExpDecay` say that its learning rate will be decayed every 1000 steps. -It is then applied like any optimser. +Here we apply exponential decay to the `Descent` optimiser. The defaults of `ExpDecay` say that its learning rate will be decayed every 1000 steps. +It is then applied like any optimiser. ```julia w = randn(10, 10) diff --git a/docs/src/training/training.md b/docs/src/training/training.md index b42db7c9..36da0eb0 100644 --- a/docs/src/training/training.md +++ b/docs/src/training/training.md @@ -7,10 +7,10 @@ To actually train a model we need four things: * A collection of data points that will be provided to the objective function. * An [optimiser](optimisers.md) that will update the model parameters appropriately. -With these we can call `Flux.train!`: +With these we can call `train!`: -```julia -Flux.train!(objective, params, data, opt) +```@docs +Flux.Optimise.train! ``` There are plenty of examples in the [model zoo](https://github.com/FluxML/model-zoo). @@ -32,6 +32,7 @@ Flux.train!(loss, ps, data, opt) ``` The objective will almost always be defined in terms of some *cost function* that measures the distance of the prediction `m(x)` from the target `y`. Flux has several of these built in, like `mse` for mean squared error or `crossentropy` for cross entropy loss, but you can calculate it however you want. +For a list of all built-in loss functions, check out the [layer reference](../models/layers.md). At first glance it may seem strange that the model that we want to train is not part of the input arguments of `Flux.train!` too. However the target of the optimizer is not the model itself, but the objective function that represents the departure between modelled and observed data. In other words, the model is implicitly defined in the objective function, and there is no need to give it explicitly. Passing the objective function instead of the model and a cost function separately provides more flexibility, and the possibility of optimizing the calculations. @@ -41,6 +42,8 @@ The model to be trained must have a set of tracked parameters that are used to c Such an object contains a reference to the model's parameters, not a copy, such that after their training, the model behaves according to their updated values. +Handling all the parameters on a layer by layer basis is explained in the [Layer Helpers](../models/basics.md) section. Also, for freezing model parameters, see the [Advanced Usage Guide](../models/advanced.md). + ## Datasets The `data` argument provides a collection of data to train with (usually a set of inputs `x` and target outputs `y`). For example, here's a dummy data set with only one data point: @@ -56,7 +59,8 @@ data = [(x, y)] ```julia data = [(x, y), (x, y), (x, y)] # Or equivalently -data = Iterators.repeated((x, y), 3) +using IterTools: ncycle +data = ncycle([(x, y)], 3) ``` It's common to load the `x`s and `y`s separately. In this case you can use `zip`: @@ -67,6 +71,14 @@ ys = [rand( 10), rand( 10), rand( 10)] data = zip(xs, ys) ``` +Training data can be conveniently partitioned for mini-batch training using the [`Flux.Data.DataLoader`](@ref) type: + +```julia +X = rand(28, 28, 60000) +Y = rand(0:9, 60000) +data = DataLoader(X, Y, batchsize=128) +``` + Note that, by default, `train!` only loops over the data once (a single "epoch"). A convenient way to run multiple epochs from the REPL is provided by `@epochs`. @@ -83,6 +95,10 @@ julia> @epochs 2 Flux.train!(...) # Train for two epochs ``` +```@docs +Flux.@epochs +``` + ## Callbacks `train!` takes an additional argument, `cb`, that's used for callbacks so that you can observe the training process. For example: @@ -120,7 +136,7 @@ An example follows that works similar to the default `Flux.train` but with no ca You don't need callbacks if you just code the calls to your functions directly into the loop. E.g. in the places marked with comments. -``` +```julia function my_custom_train!(loss, ps, data, opt) ps = Params(ps) for d in data diff --git a/docs/src/utilities.md b/docs/src/utilities.md new file mode 100644 index 00000000..7986ec23 --- /dev/null +++ b/docs/src/utilities.md @@ -0,0 +1,49 @@ +# Utility Functions + +Flux contains some utility functions for working with data; these functions +help create inputs for your models or batch your dataset. +Other functions can be used to initialize your layers or to regularly execute +callback functions. + +## Working with Data + +```@docs +Flux.unsqueeze +Flux.stack +Flux.unstack +Flux.chunk +Flux.frequencies +Flux.batch +Flux.batchseq +Base.rpad(v::AbstractVector, n::Integer, p) +``` + +## Layer Initialization + +These are primarily useful if you are planning to write your own layers. +Flux initializes convolutional layers and recurrent cells with `glorot_uniform` +by default. +To change the default on an applicable layer, pass the desired function with the +`init` keyword. For example: +```jldoctest; setup = :(using Flux) +julia> conv = Conv((3, 3), 1 => 8, relu; init=Flux.glorot_normal) +Conv((3, 3), 1=>8, relu) +``` + +```@docs +Flux.glorot_uniform +Flux.glorot_normal +``` + +## Model Abstraction + +```@docs +Flux.destructure +``` + +## Callback Helpers + +```@docs +Flux.throttle +Flux.stop +``` diff --git a/src/Flux.jl b/src/Flux.jl index 9969b323..5799fe42 100644 --- a/src/Flux.jl +++ b/src/Flux.jl @@ -7,16 +7,18 @@ using Zygote, MacroTools, Juno, Reexport, Statistics, Random using MacroTools: @forward @reexport using NNlib using Zygote: Params, @adjoint, gradient, pullback, @nograd + export gradient -export Chain, Dense, Maxout, RNN, LSTM, GRU, Conv, CrossCor, ConvTranspose, MaxPool, MeanPool, +export Chain, Dense, Maxout, RNN, LSTM, GRU, SamePad, Conv, CrossCor, ConvTranspose, + GlobalMaxPool, GlobalMeanPool, MaxPool, MeanPool, flatten, DepthwiseConv, Dropout, AlphaDropout, LayerNorm, BatchNorm, InstanceNorm, GroupNorm, - SkipConnection, params, fmap, cpu, gpu, f32, f64 + SkipConnection, params, fmap, cpu, gpu, f32, f64, testmode!, trainmode! include("optimise/Optimise.jl") using .Optimise using .Optimise: @epochs -export SGD, Descent, ADAM, Momentum, Nesterov, RMSProp, +export Descent, ADAM, Momentum, Nesterov, RMSProp, ADAGrad, AdaMax, ADADelta, AMSGrad, NADAM, ADAMW, RADAM, InvDecay, ExpDecay, WeightDecay @@ -38,24 +40,13 @@ include("data/Data.jl") include("deprecations.jl") +include("cuda/cuda.jl") + function __init__() - precompiling = ccall(:jl_generating_output, Cint, ()) != 0 - - # we don't want to include the CUDA module when precompiling, - # or we could end up replacing it at run time (triggering a warning) - precompiling && return - - if !CuArrays.functional() - # nothing to do here, and either CuArrays or one of its dependencies will have warned - else - use_cuda[] = true - - # FIXME: this functionality should be conditional at run time by checking `use_cuda` - # (or even better, get moved to CuArrays.jl as much as possible) - if CuArrays.has_cudnn() - include(joinpath(@__DIR__, "cuda/cuda.jl")) - else - @warn "CuArrays.jl did not find libcudnn. Some functionality will not be available." + use_cuda[] = CuArrays.functional() # Can be overridden after load with `Flux.use_cuda[] = false` + if CuArrays.functional() + if !CuArrays.has_cudnn() + @warn "CuArrays.jl found cuda, but did not find libcudnn. Some functionality will not be available." end end end diff --git a/src/data/Data.jl b/src/data/Data.jl index d7cd0303..16a025a7 100644 --- a/src/data/Data.jl +++ b/src/data/Data.jl @@ -3,6 +3,9 @@ module Data import ..Flux import SHA +using Random: shuffle! +using Base: @propagate_inbounds + export CMUDict, cmudict deps(path...) = joinpath(@__DIR__, "..", "..", "deps", path...) @@ -26,6 +29,9 @@ function __init__() mkpath(deps()) end +include("dataloader.jl") +export DataLoader + include("mnist.jl") export MNIST @@ -42,4 +48,7 @@ using .Sentiment include("iris.jl") export Iris +include("housing.jl") +export Housing + end diff --git a/src/data/cmudict.jl b/src/data/cmudict.jl index e6266540..9ddecbcd 100644 --- a/src/data/cmudict.jl +++ b/src/data/cmudict.jl @@ -24,18 +24,35 @@ function load() end end +""" + phones() + +Return a `Vector` containing the phones used in the CMU Pronouncing Dictionary. +""" function phones() load() Symbol.(first.(split.(split(read(deps("cmudict", "cmudict.phones"),String), "\n", keepempty = false), "\t"))) end +""" + symbols() + +Return a `Vector` containing the symbols used in the CMU Pronouncing Dictionary. +A symbol is a phone with optional auxiliary symbols, indicating for example the +amount of stress on the phone. +""" function symbols() load() Symbol.(split(read(deps("cmudict", "cmudict.symbols"),String), "\n", keepempty = false)) end +""" + rawdict() + +Return the unfiltered CMU Pronouncing Dictionary. +""" function rawdict() load() Dict(String(xs[1]) => Symbol.(xs[2:end]) for xs in @@ -44,6 +61,14 @@ end validword(s) = isascii(s) && occursin(r"^[\w\-\.]+$", s) +""" + cmudict() + +Return a filtered CMU Pronouncing Dictionary. + +It is filtered so each word contains only ASCII characters and a combination of +word characters (as determined by the regex engine using `\\w`), '-' and '.'. +""" cmudict() = filter(p -> validword(p.first), rawdict()) alphabet() = ['A':'Z'..., '0':'9'..., '_', '-', '.'] diff --git a/src/data/dataloader.jl b/src/data/dataloader.jl new file mode 100644 index 00000000..07c8f1fd --- /dev/null +++ b/src/data/dataloader.jl @@ -0,0 +1,92 @@ +# Adapted from Knet's src/data.jl (author: Deniz Yuret) + +struct DataLoader + data + batchsize::Int + nobs::Int + partial::Bool + imax::Int + indices::Vector{Int} + shuffle::Bool +end + +""" + 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 +supervised learning. The last dimension in each tensor is considered to be the observation +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. + +Example usage: + + Xtrain = rand(10, 100) + train_loader = DataLoader(Xtrain, batchsize=2) + # iterate over 50 mini-batches of size 2 + for x in train_loader + @assert size(x) == (10, 2) + ... + end + + train_loader.data # original dataset + + Xtrain = rand(10, 100) + Ytrain = rand(100) + 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) + @assert size(y) == (2,) + ... + end + end + + # train for 10 epochs + using IterTools: ncycle + Flux.train!(loss, ps, ncycle(train_loader, 10), opt) +""" +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] + for i=2:length(data) + nx != size(data[i])[end] && throw(DimensionMismatch("All data should contain same number of observations")) + end + if nx < batchsize + @warn "Number of data points less than batchsize, decreasing the batchsize to $nx" + 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 + shuffle!(d.indices) + end + nexti = min(i + d.batchsize, d.nobs) + ids = d.indices[i+1:nexti] + if length(d.data) == 1 + batch = getdata(d.data[1], ids) + else + batch = ((getdata(x, ids) for x in d.data)...,) + end + return (batch, nexti) +end + +function Base.length(d::DataLoader) + n = d.nobs / d.batchsize + d.partial ? ceil(Int,n) : floor(Int,n) +end diff --git a/src/data/fashion-mnist.jl b/src/data/fashion-mnist.jl index da78b605..5eaa1b29 100644 --- a/src/data/fashion-mnist.jl +++ b/src/data/fashion-mnist.jl @@ -33,9 +33,10 @@ const TESTLABELS = joinpath(dir, "t10k-labels-idx1-ubyte") Load the Fashion-MNIST images. -Each image is a 28×28 array of `Gray` colour values (see Colors.jl). +Each image is a 28×28 array of `Gray` colour values +(see [Colors.jl](https://github.com/JuliaGraphics/Colors.jl)). -Returns the 60,000 training images by default; pass `:test` to retreive the +Return the 60,000 training images by default; pass `:test` to retrieve the 10,000 test images. """ function images(set = :train) @@ -49,10 +50,10 @@ end labels() labels(:test) -Load the labels corresponding to each of the images returned from `images()`. +Load the labels corresponding to each of the images returned from [`images()`](@ref). Each label is a number from 0-9. -Returns the 60,000 training labels by default; pass `:test` to retreive the +Return the 60,000 training labels by default; pass `:test` to retrieve the 10,000 test labels. """ function labels(set = :train) diff --git a/src/data/housing.jl b/src/data/housing.jl new file mode 100644 index 00000000..06ecc339 --- /dev/null +++ b/src/data/housing.jl @@ -0,0 +1,136 @@ +""" +1. Title: Boston Housing Data + +2. Sources: + (a) Origin: This dataset was taken from the StatLib library which is + maintained at Carnegie Mellon University. + (b) Creator: Harrison, D. and Rubinfeld, D.L. 'Hedonic prices and the + demand for clean air', J. Environ. Economics & Management, + vol.5, 81-102, 1978. + (c) Date: July 7, 1993 + +3. Number of Instances: 506 + +4. Number of Attributes: 13 continuous attributes (including "class" + attribute "MEDV"), 1 binary-valued attribute. + +5. Attribute Information: + + 1. CRIM per capita crime rate by town + 2. ZN proportion of residential land zoned for lots over + 25,000 sq.ft. + 3. INDUS proportion of non-retail business acres per town + 4. CHAS Charles River dummy variable (= 1 if tract bounds + river; 0 otherwise) + 5. NOX nitric oxides concentration (parts per 10 million) + 6. RM average number of rooms per dwelling + 7. AGE proportion of owner-occupied units built prior to 1940 + 8. DIS weighted distances to five Boston employment centres + 9. RAD index of accessibility to radial highways + 10. TAX full-value property-tax rate per 10,000 dollars + 11. PTRATIO pupil-teacher ratio by town + 12. B 1000(Bk - 0.63)^2 where Bk is the proportion of blacks + by town + 13. LSTAT % lower status of the population + 14. MEDV Median value of owner-occupied homes in 1000's of dollars + + Downloaded From: https://archive.ics.uci.edu/ml/machine-learning-databases/housing/housing.data + +""" +module Housing + +using DelimitedFiles +using ..Data: deps, download_and_verify + +#Uncomment if package exists +#const cache_prefix = "https://cache.julialang.org/" +const cache_prefix = "" + +function load() + isfile(deps("housing.data")) && return + + @info "Downloading the Boston housing Dataset" + download_and_verify("$(cache_prefix)http://archive.ics.uci.edu/ml/machine-learning-databases/housing/housing.data", + deps("housing.data"), + "baadf72995725d76efe787b664e1f083388c79ba21ef9a7990d87f774184735a") + + #@info "Download complete. Working on the files" + path = deps() + isfile(deps("housing.data")) && touch(joinpath(path, "tempfile.data")) + open(joinpath(path, "tempfile.data"), "a") do fout + open(deps("housing.data"), "r") do fin + for line in eachline(fin) + line = replace(lstrip(line), r" +" => s",") + println(fout, line) + end + end + end + mv(joinpath(path, "tempfile.data"), deps("housing.data"), force=true) +end + +""" +Gets the targets for the Boston housing dataset, a 506 element array listing the targets for each example + +```jldoctest +julia> using Flux + +julia> target = Flux.Data.Housing.targets() + +julia> summary(target) +506×1 Array{Float64,2} + +julia> target[1] +24.0 + +""" +function targets() + load() + housing = readdlm(deps("housing.data"), ',') + reshape(Vector{Float64}(housing[1:end,end]), (506, 1)) +end + + +""" +Gets the names of the features provided in the dataset + +""" +function feature_names() + ["crim","zn","indus","chas","nox","rm","age","dis","rad","tax","ptratio","b","lstat"] +end + + +""" +Gets the features of the Boston Housing Dataset. This is a 506x13 Matrix of Float64 datatypes. +The values are in the order ["crim","zn","indus","chas","nox","rm","age","dis","rad","tax","ptratio","b","lstat"]. +It has 506 examples. + +```jldoctest +julia> using Flux + +julia> features = Flux.Data.Housing.features() + +julia> summary(features) +506×13 Array{Float64,2} + +julia> features[1, :] +13-element Array{Float64,1}: +0.00632 +18.0 +2.31 +0.0 +0.538 + ⋮ +296.0 +15.3 +396.9 +4.98 + +""" +function features() + load() + housing = readdlm(deps("housing.data"), ',') + Matrix{Float64}(housing[1:end, 1:13]) +end + + +end diff --git a/src/data/iris.jl b/src/data/iris.jl index d78606d8..76609677 100644 --- a/src/data/iris.jl +++ b/src/data/iris.jl @@ -2,13 +2,12 @@ Fisher's classic iris dataset. Measurements from 3 different species of iris: setosa, versicolor and -virginica. There are 50 examples of each species. +virginica. There are 50 examples of each species. -There are 4 measurements for each example: sepal length, sepal width, petal -length and petal width. The measurements are in centimeters. +There are 4 measurements for each example: sepal length, sepal width, +petal length and petal width. The measurements are in centimeters. The module retrieves the data from the [UCI Machine Learning Repository](https://archive.ics.uci.edu/ml/datasets/iris). - """ module Iris @@ -28,15 +27,12 @@ function load() end """ - labels() Get the labels of the iris dataset, a 150 element array of strings listing the species of each example. -```jldoctest -julia> using Flux - +```jldoctest; setup = :(Flux.Data.Iris.load()) julia> labels = Flux.Data.Iris.labels(); julia> summary(labels) @@ -53,16 +49,13 @@ function labels() end """ - features() -Get the features of the iris dataset. This is a 4x150 matrix of Float64 -elements. It has a row for each feature (sepal length, sepal width, +Get the features of the iris dataset. This is a 4x150 matrix of Float64 +elements. It has a row for each feature (sepal length, sepal width, petal length, petal width) and a column for each example. -```jldoctest -julia> using Flux - +```jldoctest; setup = :(Flux.Data.Iris.load()) julia> features = Flux.Data.Iris.features(); julia> summary(features) diff --git a/src/data/mnist.jl b/src/data/mnist.jl index b9c0540a..909814e0 100644 --- a/src/data/mnist.jl +++ b/src/data/mnist.jl @@ -83,9 +83,10 @@ getfeatures(io::IO, index::Integer) = vec(getimage(io, index)) Load the MNIST images. -Each image is a 28×28 array of `Gray` colour values (see Colors.jl). +Each image is a 28×28 array of `Gray` colour values +(see [Colors.jl](https://github.com/JuliaGraphics/Colors.jl)). -Returns the 60,000 training images by default; pass `:test` to retreive the +Return the 60,000 training images by default; pass `:test` to retrieve the 10,000 test images. """ function images(set = :train) @@ -99,10 +100,10 @@ end labels() labels(:test) -Load the labels corresponding to each of the images returned from `images()`. +Load the labels corresponding to each of the images returned from [`images()`](@ref). Each label is a number from 0-9. -Returns the 60,000 training labels by default; pass `:test` to retreive the +Return the 60,000 training labels by default; pass `:test` to retrieve the 10,000 test labels. """ function labels(set = :train) diff --git a/src/data/sentiment.jl b/src/data/sentiment.jl index ecb1ab8d..058dcf07 100644 --- a/src/data/sentiment.jl +++ b/src/data/sentiment.jl @@ -1,3 +1,4 @@ +"Stanford Sentiment Treebank dataset." module Sentiment using ZipFile @@ -39,8 +40,28 @@ function gettrees(name) return parsetree.(ss) end +""" + train() + +Return the train split of the Stanford Sentiment Treebank. +The data is in [treebank](https://en.wikipedia.org/wiki/Treebank) format. +""" train() = gettrees("train") + +""" + test() + +Return the test split of the Stanford Sentiment Treebank. +The data is in [treebank](https://en.wikipedia.org/wiki/Treebank) format. +""" test() = gettrees("test") + +""" + dev() + +Return the dev split of the Stanford Sentiment Treebank. +The data is in [treebank](https://en.wikipedia.org/wiki/Treebank) format. +""" dev() = gettrees("dev") end diff --git a/src/functor.jl b/src/functor.jl index a36b5765..0d7c55f1 100644 --- a/src/functor.jl +++ b/src/functor.jl @@ -39,6 +39,38 @@ end trainable(m) = functor(m)[1] +""" + testmode!(m, mode = true) + +Set a layer or model's test mode (see below). +Using `:auto` mode will treat any gradient computation as training. + +_Note_: if you manually set a model into test mode, you need to manually place +it back into train mode during training phase. + +Possible values include: +- `false` for training +- `true` for testing +- `:auto` or `nothing` for Flux to detect the mode automatically +""" +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)). + +_Note_: if you manually set a model into train mode, you need to manually place +it into test mode during testing phase. + +Possible values include: +- `true` for training +- `false` for testing +- `:auto` or `nothing` for Flux to detect the mode automatically +""" +trainmode!(m, mode = true) = mode isa Bool ? testmode!(m, !mode) : testmode!(m, mode) + params!(p::Params, x::AbstractArray{<:Number}, seen = IdSet()) = push!(p, x) function params!(p::Params, x, seen = IdSet()) diff --git a/src/layers/basic.jl b/src/layers/basic.jl index 6f056429..905844d7 100644 --- a/src/layers/basic.jl +++ b/src/layers/basic.jl @@ -4,17 +4,23 @@ Chain multiple layers / functions together, so that they are called in sequence on a given input. -```julia -m = Chain(x -> x^2, x -> x+1) -m(5) == 26 - -m = Chain(Dense(10, 5), Dense(5, 2)) -x = rand(10) -m(x) == m[2](m[1](x)) -``` - `Chain` also supports indexing and slicing, e.g. `m[2]` or `m[1:end-1]`. `m[1:3](x)` will calculate the output of the first three layers. + +# Examples +```jldoctest +julia> m = Chain(x -> x^2, x -> x+1); + +julia> m(5) == 26 +true + +julia> m = Chain(Dense(10, 5), Dense(5, 2)); + +julia> x = rand(10); + +julia> m(x) == m[2](m[1](x)) +true +``` """ struct Chain{T<:Tuple} layers::T @@ -33,6 +39,8 @@ applychain(fs::Tuple, x) = applychain(tail(fs), first(fs)(x)) Base.getindex(c::Chain, i::AbstractArray) = Chain(c.layers[i]...) +testmode!(m::Chain, mode = true) = (map(x -> testmode!(x, mode), m.layers); m) + function Base.show(io::IO, c::Chain) print(io, "Chain(") join(io, c.layers, ", ") @@ -58,6 +66,7 @@ outdims(c::Chain, isize) = foldl(∘, map(l -> (x -> outdims(l, x)), c.layers))( # only slightly changed to better handle interaction with Zygote @dsweber2 """ activations(c::Chain, input) + Calculate the forward results of each layers in Chain `c` with `input` as model input. """ function activations(c::Chain, input) @@ -76,22 +85,22 @@ extraChain(::Tuple{}, x) = () """ Dense(in::Integer, out::Integer, σ = identity) -Creates a traditional `Dense` layer with parameters `W` and `b`. +Create a traditional `Dense` layer with parameters `W` and `b`. y = σ.(W * x .+ b) The input `x` must be a vector of length `in`, or a batch of vectors represented as an `in × N` matrix. The out `y` will be a vector or batch of length `out`. -```julia +# Examples +```jldoctest; setup = :(using Random; Random.seed!(0)) julia> d = Dense(5, 2) Dense(5, 2) julia> d(rand(5)) -Tracked 2-element Array{Float64,1}: - 0.00257447 - -0.00449443 -``` +2-element Array{Float32,1}: + -0.16210233 + 0.12311903``` """ struct Dense{F,S,T} W::S @@ -143,7 +152,7 @@ outdims(l::Dense, isize) = (size(l.W)[1],) """ Diagonal(in::Integer) -Creates an element-wise linear transformation layer with learnable +Create an element-wise linear transformation layer with learnable vectors `α` and `β`: y = α .* x .+ β @@ -174,18 +183,11 @@ outdims(l::Diagonal, isize) = (length(l.α),) """ Maxout(over) -`Maxout` is a neural network layer, which has a number of internal layers, -which all have the same input, and the maxout returns the elementwise maximium -of the internal layers' outputs. +The [Maxout](https://arxiv.org/pdf/1302.4389.pdf) layer has a number of +internal layers which all receive the same input. It returns the elementwise +maximum of the internal layers' outputs. Maxout over linear dense layers satisfies the univeral approximation theorem. - -Reference: -Ian J. Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron Courville, and Yoshua Bengio. -2013. Maxout networks. -In Proceedings of the 30th International Conference on International Conference on Machine Learning - Volume 28 (ICML'13), -Sanjoy Dasgupta and David McAllester (Eds.), Vol. 28. JMLR.org III-1319-III-1327. -https://arxiv.org/pdf/1302.4389.pdf """ struct Maxout{FS<:Tuple} over::FS @@ -194,17 +196,18 @@ end """ Maxout(f, n_alts) -Constructs a Maxout layer over `n_alts` instances of the layer given by `f`. -The function takes no arguement and should return some callable layer. -Conventionally this is a linear dense layer. +Construct a Maxout layer over `n_alts` instances of the layer given by `f`. +The function takes no arguments and should return some callable layer. +Conventionally, this is a linear dense layer. -For example the following example which -will construct a `Maxout` layer over 4 internal dense linear layers, -each identical in structure (784 inputs, 128 outputs). +# Examples + +This constructs a `Maxout` layer over 4 internal dense linear layers, each +identical in structure (784 inputs, 128 outputs): ```julia - insize = 784 - outsize = 128 - Maxout(()->Dense(insize, outsize), 4) +insize = 784 +outsize = 128 +Maxout(()->Dense(insize, outsize), 4) ``` """ function Maxout(f, n_alts) @@ -221,16 +224,18 @@ end outdims(l::Maxout, isize) = outdims(first(l.over), isize) """ - SkipConnection(layers, connection) + SkipConnection(layer, connection) -Creates a Skip Connection, of a layer or `Chain` of consecutive layers -plus a shortcut connection. The connection function will combine the result of the layers -with the original input, to give the final output. +Create a skip connection which consists of a layer or `Chain` of consecutive +layers and a shortcut connection linking the block's input to the output +through a user-supplied 2-argument callable. The first argument to the callable +will be propagated through the given `layer` while the second is the unchanged, +"skipped" input. -The simplest 'ResNet'-type connection is just `SkipConnection(layer, +)`, +The simplest "ResNet"-type connection is just `SkipConnection(layer, +)`, and requires the output of the layers to be the same shape as the input. Here is a more complicated example: -``` +```julia m = Conv((3,3), 4=>7, pad=(1,1)) x = ones(5,5,4,10); size(m(x)) == (5, 5, 7, 10) diff --git a/src/layers/conv.jl b/src/layers/conv.jl index 997b96e5..1c7282b5 100644 --- a/src/layers/conv.jl +++ b/src/layers/conv.jl @@ -9,20 +9,38 @@ expand(N, i::Tuple) = i expand(N, i::Integer) = ntuple(_ -> i, N) """ - Conv(filter::Tuple, in=>out) - Conv(filter::Tuple, in=>out, activation) + SamePad -Standard convolutional layer. `filter` should be a tuple like `(2, 2)`. -`in` and `out` specify the number of input and output channels respectively. +Padding for convolutional layers will be calculated so that outputshape == inputshape when stride = 1. -Example: Applying Conv layer to a 1-channel input using a 2x2 window size, - giving us a 16-channel output. Output is activated with ReLU. +For stride > 1 the output shape depends on the type of convolution layer. +""" +struct SamePad end + +calc_padding(pad, k::NTuple{N,T}, dilation, stride) where {T,N}= expand(Val(2*N), pad) +function calc_padding(::SamePad, k::NTuple{N,T}, dilation, stride) where {N,T} + #Ref: "A guide to convolution arithmetic for deep learning" https://arxiv.org/pdf/1603.07285 + + # Effective kernel size, including dilation + k_eff = @. k + (k - 1) * (dilation - 1) + # How much total padding needs to be applied? + pad_amt = @. k_eff - 1 + # In case amount of padding is odd we need to apply different amounts to each side. + return Tuple(mapfoldl(i -> [ceil(Int, i/2), floor(Int, i/2)], vcat, pad_amt)) +end + +""" + Conv(filter, in => out, σ = identity; init = glorot_uniform, + stride = 1, pad = 0, dilation = 1) filter = (2,2) in = 1 out = 16 Conv((2, 2), 1=>16, relu) +Standard convolutional layer. `filter` should be a tuple like `(2, 2)`. +`in` and `out` specify the number of input and output channels respectively. + Data should be stored in WHCN order (width, height, # channels, batch size). In other words, a 100×100 RGB image would be a `100×100×3×1` array, and a batch of 50 would be a `100×100×3×50` array. @@ -31,6 +49,18 @@ Accepts keyword arguments `weight` and `bias` to set the corresponding fields. Setting `bias` to `Flux.Zeros()` will switch bias off for the layer. Takes the keyword arguments `pad`, `stride` and `dilation`. +Use `pad=SamePad()` to apply padding so that outputsize == inputsize / stride. + +# Examples + +Apply a `Conv` layer to a 1-channel input using a 2×2 window filter size, giving us a +16-channel output. Output is activated with ReLU. +```julia +filter = (2,2) +in = 1 +out = 16 +Conv(filter, in => out, relu) +``` """ struct Conv{N,M,F,A,V} σ::F @@ -46,18 +76,27 @@ end Conv(weight::AbstractArray, bias::AbstractArray, activation) Constructs the convolutional layer with user defined weight and bias arrays. -All other behaviours of the Conv layer apply with regard to data order and -forward pass. Setting `bias` to `Flux.Zeros()` would switch `bias` off for the layer. Takes the keyword arguments `pad`, `stride` and `dilation`. + +There is also a keyword-only constuctor available for all convoultional +layers. + +```julia +weight = rand(Float32, 3, 3, 5) +bias = zeros(Float32, 5) +Conv(weight = weight, + bias = bias, + σ = sigmoid) +``` """ function Conv(w::AbstractArray{T,N}, b::Union{Zeros, AbstractVector{T}}, σ = identity; stride = 1, pad = 0, dilation = 1) where {T,N} stride = expand(Val(N-2), stride) - pad = expand(Val(2*(N-2)), pad) dilation = expand(Val(N-2), dilation) + pad = calc_padding(pad, size(w)[1:N-2], dilation, stride) return Conv(σ, w, b, stride, pad, dilation) end @@ -114,8 +153,8 @@ end """ outdims(l::Conv, isize::Tuple) -Calculate the output dimensions given the input dimensions, `isize`. -Batch size and channel size are ignored as per `NNlib.jl`. +Calculate the output dimensions given the input dimensions `isize`. +Batch size and channel size are ignored as per [NNlib.jl](https://github.com/FluxML/NNlib.jl). ```julia m = Conv((3, 3), 3 => 16) @@ -127,19 +166,23 @@ outdims(l::Conv, isize) = output_size(DenseConvDims(_paddims(isize, size(l.weight)), size(l.weight); stride = l.stride, padding = l.pad, dilation = l.dilation)) """ - ConvTranspose(size, in=>out) - ConvTranspose(size, in=>out, activation) + ConvTranspose(filter, in=>out) + ConvTranspose(filter, in=>out, activation) + ConvTranspose(filter, in => out, σ = identity; init = glorot_uniform, + stride = 1, pad = 0, dilation = 1) Standard convolutional transpose layer. `filter` should be a tuple like `(2, 2)`. `in` and `out` specify the number of input and output channels respectively. -Data should be stored in WHCN order. In other words, a 100×100 RGB image would -be a `100×100×3` array, and a batch of 50 would be a `100×100×3×50` array. +Data should be stored in WHCN order (width, height, # channels, batch size). +In other words, a 100×100 RGB image would be a `100×100×3×1` array, +and a batch of 50 would be a `100×100×3×50` array. Accepts keyword arguments `weight` and `bias` to set the corresponding fields. Setting `bias` to `Flux.Zeros()` will switch bias off for the layer. Takes the keyword arguments `pad`, `stride` and `dilation`. +Use `pad=SamePad()` to apply padding so that outputsize == stride * inputsize - stride + 1. """ struct ConvTranspose{N,M,F,A,V} σ::F @@ -155,18 +198,19 @@ end ConvTranspose(weight::AbstractArray, bias::AbstractArray, activation) Constructs the convolutional transpose layer with user defined weight and bias arrays. -All other behaviours of the ConvTranspose layer apply with regard to data order and forward pass. Setting `bias` to `Flux.Zeros()` would switch `bias` off for the layer. Takes the keyword arguments `pad`, `stride` and `dilation`. + +For keyword-only constuctor, see also [`Conv`](@ref) """ function ConvTranspose(w::AbstractArray{T,N}, b::Union{Zeros, AbstractVector{T}}, σ = identity; stride = 1, pad = 0, dilation = 1) where {T,N} stride = expand(Val(N-2), stride) - pad = expand(Val(2*(N-2)), pad) dilation = expand(Val(N-2), dilation) + pad = calc_padding(pad, size(w)[1:N-2], dilation, stride) return ConvTranspose(σ, w, b, stride, pad, dilation) end @@ -227,18 +271,22 @@ outdims(l::ConvTranspose{N}, isize) where N = _convtransoutdims(isize[1:2], size """ DepthwiseConv(filter::Tuple, in=>out) DepthwiseConv(filter::Tuple, in=>out, activation) + DepthwiseConv(filter, in => out, σ = identity; init = glorot_uniform, + stride = 1, pad = 0, dilation = 1) Depthwise convolutional layer. `filter` should be a tuple like `(2, 2)`. `in` and `out` specify the number of input and output channels respectively. Note that `out` must be an integer multiple of `in`. -Data should be stored in WHCN order. In other words, a 100×100 RGB image would -be a `100×100×3` array, and a batch of 50 would be a `100×100×3×50` array. +Data should be stored in WHCN order (width, height, # channels, batch size). +In other words, a 100×100 RGB image would be a `100×100×3×1` array, +and a batch of 50 would be a `100×100×3×50` array. Accepts keyword arguments `weight` and `bias` to set the corresponding fields. Setting `bias` to `Flux.Zeros()` will switch bias off for the layer. Takes the keyword arguments `pad`, `stride` and `dilation`. +Use `pad=SamePad()` to apply padding so that outputsize == inputsize / stride. """ struct DepthwiseConv{N,M,F,A,V} σ::F @@ -254,18 +302,19 @@ end DepthwiseConv(weight::AbstractArray, bias::AbstractArray, activation) Constructs the `DepthwiseConv` layer with user defined weight and bias arrays. -All other behaviours of the `DepthwiseConv` layer apply with regard to data order and forward pass. Setting `bias` to `Flux.Zeros()` would switch `bias` off for the layer. Takes the keyword arguments `pad`, `stride` and `dilation`. + +For keyword-only constuctor, see also [`Conv`](@ref) """ function DepthwiseConv(w::AbstractArray{T,N}, b::Union{Zeros, AbstractVector{T}}, σ = identity; stride = 1, pad = 0, dilation = 1) where {T,N} stride = expand(Val(N-2), stride) - pad = expand(Val(2*(N-2)), pad) dilation = expand(Val(N-2), dilation) + pad = calc_padding(pad, size(w)[1:N-2], dilation, stride) return DepthwiseConv(σ, w, b, stride, pad, dilation) end @@ -328,21 +377,15 @@ outdims(l::DepthwiseConv, isize) = output_size(DepthwiseConvDims(_paddims(isize, (1, 1, size(l.weight)[end], 1)), size(l.weight); stride = l.stride, padding = l.pad, dilation = l.dilation)) """ - CrossCor(size, in=>out) - CrossCor(size, in=>out, activation) + CrossCor(filter, in=>out) + CrossCor(filter, in=>out, activation) + CrossCor(filter, in => out, σ = identity; init = glorot_uniform, + stride = 1, pad = 0, dilation = 1) -Standard cross convolutional layer. `size` should be a tuple like `(2, 2)`. +Standard cross convolutional layer. `filter` should be a tuple like `(2, 2)`. `in` and `out` specify the number of input and output channels respectively. -Example: Applying CrossCor layer to a 1-channel input using a 2x2 window size, - giving us a 16-channel output. Output is activated with ReLU. - - size = (2,2) - in = 1 - out = 16 - CrossCor((2, 2), 1=>16, relu) - -Data should be stored in WHCN order (width, height, # channels, # batches). +Data should be stored in WHCN order (width, height, # channels, batch size). In other words, a 100×100 RGB image would be a `100×100×3×1` array, and a batch of 50 would be a `100×100×3×50` array. @@ -350,6 +393,18 @@ Accepts keyword arguments `weight` and `bias` to set the corresponding fields. Setting `bias` to `Flux.Zeros()` will switch bias off for the layer. Takes the keyword arguments `pad`, `stride` and `dilation`. +Use `pad=SamePad()` to apply padding so that outputsize == inputsize / stride. + +# Examples + +Apply a `CrossCor` layer to a 1-channel input using a 2×2 window filter size, giving us a +16-channel output. Output is activated with ReLU. +```julia +filter = (2,2) +in = 1 +out = 16 +CrossCor((2, 2), 1=>16, relu) +``` """ struct CrossCor{N,M,F,A,V} σ::F @@ -365,18 +420,19 @@ end CrossCor(weight::AbstractArray, bias::AbstractArray, activation) Constructs the standard cross convolutional layer with user defined weight and bias -arrays. All other behaviours of the CrossCor layer apply with regard to data order and -forward pass. +arrays. Setting `bias` to `Flux.Zeros()` would switch `bias` off for the layer. Takes the keyword arguments `pad`, `stride` and `dilation`. + +For keyword-only constuctor, see also [`Conv`](@ref) """ function CrossCor(w::AbstractArray{T,N}, b::Union{Zeros, AbstractVector{T}}, σ = identity; stride = 1, pad = 0, dilation = 1) where {T,N} stride = expand(Val(N-2), stride) - pad = expand(Val(2*(N-2)), pad) dilation = expand(Val(N-2), dilation) + pad = calc_padding(pad, size(w)[1:N-2], dilation, stride) return CrossCor(σ, w, b, stride, pad, dilation) end @@ -425,11 +481,62 @@ outdims(l::CrossCor, isize) = output_size(DenseConvDims(_paddims(isize, size(l.weight)), size(l.weight); stride = l.stride, padding = l.pad, dilation = l.dilation)) """ - MaxPool(k) + GlobalMaxPool() -Max pooling layer. `k` stands for the size of the window for each dimension of the input. +Global max pooling layer. -Takes the keyword arguments `pad` and `stride`. +Transforms (w,h,c,b)-shaped input into (1,1,c,b)-shaped output, +by performing max pooling on the complete (w,h)-shaped feature maps. +""" +struct GlobalMaxPool end + +function (g::GlobalMaxPool)(x) + # Input size + x_size = size(x) + # Kernel size + k = x_size[1:end-2] + # Pooling dimensions + pdims = PoolDims(x, k) + + return maxpool(x, pdims) +end + +function Base.show(io::IO, g::GlobalMaxPool) + print(io, "GlobalMaxPool()") +end + +""" + GlobalMeanPool() + +Global mean pooling layer. + +Transforms (w,h,c,b)-shaped input into (1,1,c,b)-shaped output, +by performing mean pooling on the complete (w,h)-shaped feature maps. +""" +struct GlobalMeanPool end + +function (g::GlobalMeanPool)(x) + # Input size + x_size = size(x) + # Kernel size + k = x_size[1:end-2] + # Pooling dimensions + pdims = PoolDims(x, k) + + return meanpool(x, pdims) +end + +function Base.show(io::IO, g::GlobalMeanPool) + print(io, "GlobalMeanPool()") +end + +""" + MaxPool(k; pad = 0, stride = k) + +Max pooling layer. `k` is the size of the window for each dimension of the input. + +Use `pad=SamePad()` to apply padding so that outputsize == inputsize / stride. +======= """ struct MaxPool{N,M} k::NTuple{N,Int} @@ -439,8 +546,7 @@ end function MaxPool(k::NTuple{N,Integer}; pad = 0, stride = k) where N stride = expand(Val(N), stride) - pad = expand(Val(2*N), pad) - + pad = calc_padding(pad, k, 1, stride) return MaxPool(k, pad, stride) end @@ -456,11 +562,11 @@ end outdims(l::MaxPool{N}, isize) where N = output_size(PoolDims(_paddims(isize, (l.k..., 1, 1)), l.k; stride = l.stride, padding = l.pad)) """ - MeanPool(k) + MeanPool(k; pad = 0, stride = k) -Mean pooling layer. `k` stands for the size of the window for each dimension of the input. +Mean pooling layer. `k` is the size of the window for each dimension of the input. -Takes the keyword arguments `pad` and `stride`. +Use `pad=SamePad()` to apply padding so that outputsize == inputsize / stride. """ struct MeanPool{N,M} k::NTuple{N,Int} @@ -470,7 +576,7 @@ end function MeanPool(k::NTuple{N,Integer}; pad = 0, stride = k) where N stride = expand(Val(N), stride) - pad = expand(Val(2*N), pad) + pad = calc_padding(pad, k, 1, stride) return MeanPool(k, pad, stride) end @@ -483,4 +589,4 @@ function Base.show(io::IO, m::MeanPool) print(io, "MeanPool(", m.k, ", pad = ", m.pad, ", stride = ", m.stride, ")") end -outdims(l::MeanPool{N}, isize) where N = output_size(PoolDims(_paddims(isize, (l.k..., 1, 1)), l.k; stride = l.stride, padding = l.pad)) \ No newline at end of file +outdims(l::MeanPool{N}, isize) where N = output_size(PoolDims(_paddims(isize, (l.k..., 1, 1)), l.k; stride = l.stride, padding = l.pad)) diff --git a/src/layers/normalise.jl b/src/layers/normalise.jl index b421d3e7..0b5e04fb 100644 --- a/src/layers/normalise.jl +++ b/src/layers/normalise.jl @@ -2,11 +2,23 @@ istraining() = false @adjoint istraining() = true, _ -> nothing +_isactive(m) = isnothing(m.active) ? istraining() : m.active + _dropout_shape(s, ::Colon) = size(s) _dropout_shape(s, dims) = tuple((i ∉ dims ? 1 : si for (i, si) ∈ enumerate(size(s)))...) _dropout_kernel(y::T, p, q) where {T} = y > p ? T(1 / q) : T(0) +""" + dropout(x, p; dims = :) + +The dropout function. For each input, either sets that input to `0` (with probability +`p`) or scales it by `1 / (1 - p)`. `dims` specifies the unbroadcasted dimensions, +e.g. `dims=1` applies dropout along columns and `dims=2` along rows. +This is used as a regularisation, i.e. it reduces overfitting during training. + +See also the [`Dropout`](@ref) layer. +""" dropout(x, p; dims = :) = x @adjoint function dropout(x, p; dims = :) @@ -18,22 +30,31 @@ end """ Dropout(p, dims = :) -A Dropout layer. For each input, either sets that input to `0` (with probability -`p`) or scales it by `1/(1-p)`. The `dims` argument is to specified the unbroadcasted - dimensions, i.e. `dims=1` does dropout along columns and `dims=2` along rows. This is - used as a regularisation, i.e. it reduces overfitting during training. see also [`dropout`](@ref). +Dropout layer. In the forward pass, apply the [`Flux.dropout`](@ref) function on the input. + +Does nothing to the input once [`Flux.testmode!`](@ref) is `true`. """ mutable struct Dropout{F,D} p::F dims::D + active::Union{Bool, Nothing} end +# TODO: deprecate in v0.11 +Dropout(p, dims) = Dropout(p, dims, nothing) + function Dropout(p; dims = :) @assert 0 ≤ p ≤ 1 - Dropout{typeof(p),typeof(dims)}(p, dims) + Dropout{typeof(p),typeof(dims)}(p, dims, nothing) end -(a::Dropout)(x) = dropout(x, a.p; dims = a.dims) +function (a::Dropout)(x) + _isactive(a) || return x + return dropout(x, a.p; dims = a.dims) +end + +testmode!(m::Dropout, mode = true) = + (m.active = (isnothing(mode) || mode == :auto) ? nothing : !mode; m) function Base.show(io::IO, d::Dropout) print(io, "Dropout(", d.p) @@ -43,20 +64,25 @@ end """ AlphaDropout(p) -A dropout layer. It is used in Self-Normalizing Neural Networks. -(https://papers.nips.cc/paper/6698-self-normalizing-neural-networks.pdf) -The AlphaDropout layer ensures that mean and variance of activations remains the same as before. + +A dropout layer. Used in +[Self-Normalizing Neural Networks](https://papers.nips.cc/paper/6698-self-normalizing-neural-networks.pdf). +The AlphaDropout layer ensures that mean and variance of activations +remain the same as before. + +Does nothing to the input once [`testmode!`](@ref) is true. """ mutable struct AlphaDropout{F} p::F - function AlphaDropout(p) + active::Union{Bool, Nothing} + function AlphaDropout(p, active = nothing) @assert 0 ≤ p ≤ 1 - new{typeof(p)}(p) + new{typeof(p)}(p, active) end end function (a::AlphaDropout)(x) - istraining() || return x + _isactive(a) || return x λ = eltype(x)(1.0507009873554804934193349852946) α = eltype(x)(1.6732632423543772848170429916717) α1 = eltype(x)(-λ*α) @@ -68,12 +94,15 @@ function (a::AlphaDropout)(x) return x end +testmode!(m::AlphaDropout, mode = true) = + (m.active = (isnothing(mode) || mode == :auto) ? nothing : !mode; m) + """ LayerNorm(h::Integer) A [normalisation layer](https://arxiv.org/pdf/1607.06450.pdf) designed to be -used with recurrent hidden states of size `h`. Normalises the mean/stddev of -each input before applying a per-neuron gain/bias. +used with recurrent hidden states of size `h`. Normalises the mean and standard +deviation of each input before applying a per-neuron gain/bias. """ struct LayerNorm{T} diag::Diagonal{T} @@ -95,8 +124,8 @@ end initβ = zeros, initγ = ones, ϵ = 1e-8, momentum = .1) -Batch Normalization layer. The `channels` input should be the size of the -channel dimension in your data (see below). +[Batch Normalization](https://arxiv.org/pdf/1502.03167.pdf) layer. +`channels` should be the size of the channel dimension in your data (see below). Given an array with `N` dimensions, call the `N-1`th the channel dimension. (For a batch of feature vectors this is just the data dimension, for `WHCN` images @@ -106,10 +135,9 @@ it's the usual channel dimension.) shifts them to have a new mean and variance (corresponding to the learnable, per-channel `bias` and `scale` parameters). -See [Batch Normalization: Accelerating Deep Network Training by Reducing -Internal Covariate Shift](https://arxiv.org/pdf/1502.03167.pdf). +Use [`testmode!`](@ref) during inference. -Example: +# Examples ```julia m = Chain( Dense(28^2, 64), @@ -127,12 +155,16 @@ mutable struct BatchNorm{F,V,W,N} σ²::W # moving std ϵ::N momentum::N + active::Union{Bool, Nothing} end +# TODO: deprecate in v0.11 +BatchNorm(λ, β, γ, μ, σ², ϵ, momentum) = BatchNorm(λ, β, γ, μ, σ², ϵ, momentum, nothing) + BatchNorm(chs::Integer, λ = identity; initβ = (i) -> zeros(Float32, i), initγ = (i) -> ones(Float32, i), ϵ = 1f-5, momentum = 0.1f0) = BatchNorm(λ, initβ(chs), initγ(chs), - zeros(chs), ones(chs), ϵ, momentum) + zeros(chs), ones(chs), ϵ, momentum, nothing) trainable(bn::BatchNorm) = (bn.β, bn.γ) @@ -145,7 +177,7 @@ function (BN::BatchNorm)(x) m = div(prod(size(x)), channels) γ = reshape(BN.γ, affine_shape...) β = reshape(BN.β, affine_shape...) - if !istraining() + if !_isactive(BN) μ = reshape(BN.μ, affine_shape...) σ² = reshape(BN.σ², affine_shape...) ϵ = BN.ϵ @@ -170,41 +202,15 @@ end @functor BatchNorm +testmode!(m::BatchNorm, mode = true) = + (m.active = (isnothing(mode) || mode == :auto) ? nothing : !mode; m) + function Base.show(io::IO, l::BatchNorm) print(io, "BatchNorm($(join(size(l.β), ", "))") (l.λ == identity) || print(io, ", λ = $(l.λ)") print(io, ")") end - -""" - InstanceNorm(channels::Integer, σ = identity; - initβ = zeros, initγ = ones, - ϵ = 1e-8, momentum = .1) - -Instance Normalization layer. The `channels` input should be the size of the -channel dimension in your data (see below). - -Given an array with `N` dimensions, call the `N-1`th the channel dimension. (For -a batch of feature vectors this is just the data dimension, for `WHCN` images -it's the usual channel dimension.) - -`InstanceNorm` computes the mean and variance for each each `W×H×1×1` slice and -shifts them to have a new mean and variance (corresponding to the learnable, -per-channel `bias` and `scale` parameters). - -See [Instance Normalization: The Missing Ingredient for Fast Stylization](https://arxiv.org/abs/1607.08022). - -Example: -```julia -m = Chain( - Dense(28^2, 64), - InstanceNorm(64, relu), - Dense(64, 10), - InstanceNorm(10), - softmax) -``` -""" expand_inst = (x, as) -> reshape(repeat(x, outer=[1, as[length(as)]]), as...) mutable struct InstanceNorm{F,V,W,N} @@ -215,12 +221,44 @@ mutable struct InstanceNorm{F,V,W,N} σ²::W # moving std ϵ::N momentum::N + active::Union{Bool, Nothing} end +# TODO: deprecate in v0.11 +""" + InstanceNorm(channels::Integer, σ = identity; + initβ = zeros, initγ = ones, + ϵ = 1e-8, momentum = .1) + +[Instance Normalization](https://arxiv.org/abs/1607.08022) layer. +`channels` should be the size of the channel dimension in your data (see below). + +Given an array with `N` dimensions, call the `N-1`th the channel dimension. (For +a batch of feature vectors this is just the data dimension, for `WHCN` images +it's the usual channel dimension.) + +`InstanceNorm` computes the mean and variance for each each `W×H×1×1` slice and +shifts them to have a new mean and variance (corresponding to the learnable, +per-channel `bias` and `scale` parameters). + +Use [`testmode!`](@ref) during inference. + +# Examples +```julia +m = Chain( + Dense(28^2, 64), + InstanceNorm(64, relu), + Dense(64, 10), + InstanceNorm(10), + softmax) +``` +""" +InstanceNorm(λ, β, γ, μ, σ², ϵ, momentum) = InstanceNorm(λ, β, γ, μ, σ², ϵ, momentum, nothing) + InstanceNorm(chs::Integer, λ = identity; initβ = (i) -> zeros(Float32, i), initγ = (i) -> ones(Float32, i), ϵ = 1f-5, momentum = 0.1f0) = InstanceNorm(λ, initβ(chs), initγ(chs), - zeros(chs), ones(chs), ϵ, momentum) + zeros(chs), ones(chs), ϵ, momentum, nothing) trainable(in::InstanceNorm) = (in.β, in.γ) @@ -237,7 +275,7 @@ function (in::InstanceNorm)(x) m = div(prod(size(x)), c*bs) γ, β = expand_inst(in.γ, affine_shape), expand_inst(in.β, affine_shape) - if !istraining() + if !_isactive(in) μ = expand_inst(in.μ, affine_shape) σ² = expand_inst(in.σ², affine_shape) ϵ = in.ϵ @@ -263,6 +301,9 @@ end @functor InstanceNorm +testmode!(m::InstanceNorm, mode = true) = + (m.active = (isnothing(mode) || mode == :auto) ? nothing : !mode; m) + function Base.show(io::IO, l::InstanceNorm) print(io, "InstanceNorm($(join(size(l.β), ", "))") (l.λ == identity) || print(io, ", λ = $(l.λ)") @@ -270,26 +311,27 @@ function Base.show(io::IO, l::InstanceNorm) end """ -Group Normalization. -This layer can outperform Batch-Normalization and Instance-Normalization. + GroupNorm(chs::Integer, G::Integer, λ = identity; + initβ = (i) -> zeros(Float32, i), initγ = (i) -> ones(Float32, i), + ϵ = 1f-5, momentum = 0.1f0) - GroupNorm(chs::Integer, G::Integer, λ = identity; - initβ = (i) -> zeros(Float32, i), initγ = (i) -> ones(Float32, i), - ϵ = 1f-5, momentum = 0.1f0) +[Group Normalization](https://arxiv.org/pdf/1803.08494.pdf) layer. +This layer can outperform Batch Normalization and Instance Normalization. -``chs`` is the number of channels, the channel dimension of your input. -For an array of N dimensions, the (N-1)th index is the channel dimension. +`chs` is the number of channels, the channel dimension of your input. +For an array of N dimensions, the `N-1`th index is the channel dimension. -``G`` is the number of groups along which the statistics would be computed. +`G` is the number of groups along which the statistics are computed. The number of channels must be an integer multiple of the number of groups. -Example: -``` -m = Chain(Conv((3,3), 1=>32, leakyrelu;pad = 1), - GroupNorm(32,16)) # 32 channels, 16 groups (G = 16), thus 2 channels per group used -``` +Use [`testmode!`](@ref) during inference. -Link : https://arxiv.org/pdf/1803.08494.pdf +# Examples +```julia +m = Chain(Conv((3,3), 1=>32, leakyrelu;pad = 1), + GroupNorm(32,16)) + # 32 channels, 16 groups (G = 16), thus 2 channels per group used +``` """ mutable struct GroupNorm{F,V,W,N,T} G::T # number of groups @@ -300,12 +342,16 @@ mutable struct GroupNorm{F,V,W,N,T} σ²::W # moving std ϵ::N momentum::N + active::Union{Bool, Nothing} end +# TODO: deprecate in v0.11 +GroupNorm(G, λ, β, γ, μ, σ², ϵ, momentum) = GroupNorm(G, λ, β, γ, μ, σ², ϵ, momentum, nothing) + GroupNorm(chs::Integer, G::Integer, λ = identity; initβ = (i) -> zeros(Float32, i), initγ = (i) -> ones(Float32, i), ϵ = 1f-5, momentum = 0.1f0) = GroupNorm(G, λ, initβ(chs), initγ(chs), - zeros(G,1), ones(G,1), ϵ, momentum) + zeros(G,1), ones(G,1), ϵ, momentum, nothing) trainable(gn::GroupNorm) = (gn.β, gn.γ) @@ -329,7 +375,7 @@ function(gn::GroupNorm)(x) β = reshape(gn.β, affine_shape...) y = reshape(x,((size(x))[1:end-2]...,channels_per_group,groups,batches)) - if !istraining() + if !_isactive(gn) og_shape = size(x) μ = reshape(gn.μ, μ_affine_shape...) # Shape : (1,1,...C/G,G,1) σ² = reshape(gn.σ², μ_affine_shape...) # Shape : (1,1,...C/G,G,1) @@ -360,6 +406,9 @@ end @functor GroupNorm +testmode!(m::GroupNorm, mode = true) = + (m.active = (isnothing(mode) || mode == :auto) ? nothing : !mode; m) + function Base.show(io::IO, l::GroupNorm) print(io, "GroupNorm($(join(size(l.β), ", "))") (l.λ == identity) || print(io, ", λ = $(l.λ)") diff --git a/src/layers/recurrent.jl b/src/layers/recurrent.jl index 647dda25..a93c4a0a 100644 --- a/src/layers/recurrent.jl +++ b/src/layers/recurrent.jl @@ -12,16 +12,16 @@ in the background. `cell` should be a model of the form: h, y = cell(h, x...) -For example, here's a recurrent network that keeps a running total of its inputs. +For example, here's a recurrent network that keeps a running total of its inputs: ```julia -accum(h, x) = (h+x, x) +accum(h, x) = (h + x, x) rnn = Flux.Recur(accum, 0) -rnn(2) # 2 -rnn(3) # 3 -rnn.state # 5 -rnn.(1:10) # apply to a sequence -rnn.state # 60 +rnn(2) # 2 +rnn(3) # 3 +rnn.state # 5 +rnn.(1:10) # apply to a sequence +rnn.state # 60 ``` """ mutable struct Recur{T} @@ -47,9 +47,10 @@ Base.show(io::IO, m::Recur) = print(io, "Recur(", m.cell, ")") Reset the hidden state of a recurrent layer back to its original value. -Assuming you have a `Recur` layer `rnn`, this is roughly equivalent to - - rnn.state = hidden(rnn.cell) +Assuming you have a `Recur` layer `rnn`, this is roughly equivalent to: +```julia +rnn.state = hidden(rnn.cell) +``` """ reset!(m::Recur) = (m.state = m.init) reset!(m) = foreach(reset!, functor(m)[1]) @@ -135,8 +136,8 @@ Base.show(io::IO, l::LSTMCell) = """ LSTM(in::Integer, out::Integer) -Long Short Term Memory recurrent layer. Behaves like an RNN but generally -exhibits a longer memory span over sequences. +[Long Short Term Memory](https://www.researchgate.net/publication/13853244_Long_Short-term_Memory) +recurrent layer. Behaves like an RNN but generally exhibits a longer memory span over sequences. See [this article](https://colah.github.io/posts/2015-08-Understanding-LSTMs/) for a good overview of the internals. @@ -176,8 +177,8 @@ Base.show(io::IO, l::GRUCell) = """ GRU(in::Integer, out::Integer) -Gated Recurrent Unit layer. Behaves like an RNN but generally -exhibits a longer memory span over sequences. +[Gated Recurrent Unit](https://arxiv.org/abs/1406.1078) layer. Behaves like an +RNN but generally exhibits a longer memory span over sequences. See [this article](https://colah.github.io/posts/2015-08-Understanding-LSTMs/) for a good overview of the internals. diff --git a/src/layers/stateless.jl b/src/layers/stateless.jl index 159a8385..3f97e1fd 100644 --- a/src/layers/stateless.jl +++ b/src/layers/stateless.jl @@ -1,10 +1,58 @@ -using CuArrays -using NNlib: logsoftmax, logσ - # Cost functions +""" + mae(ŷ, y) +Return the mean of absolute error; calculated as +`sum(abs.(ŷ .- y)) / length(y)`. +""" +mae(ŷ, y) = sum(abs.(ŷ .- y)) * 1 // length(y) + + +""" + mse(ŷ, y) + +Return the mean squared error between ŷ and y; calculated as +`sum((ŷ .- y).^2) / length(y)`. + +# Examples +```jldoctest +julia> Flux.mse([0, 2], [1, 1]) +1//1 +``` +""" mse(ŷ, y) = sum((ŷ .- y).^2) * 1 // length(y) + +""" + msle(ŷ, y; ϵ=eps(eltype(ŷ))) + +Return the mean of the squared logarithmic errors; calculated as +`sum((log.(ŷ .+ ϵ) .- log.(y .+ ϵ)).^2) / length(y)`. +The `ϵ` term provides numerical stability. + +Penalizes an under-predicted estimate greater than an over-predicted estimate. +""" +msle(ŷ, y; ϵ=eps(eltype(ŷ))) = sum((log.(ŷ .+ ϵ) .- log.(y .+ ϵ)).^2) * 1 // length(y) + + + +""" + huber_loss(ŷ, y; δ=1.0) + +Return the mean of the [Huber loss](https://en.wikipedia.org/wiki/Huber_loss) +given the prediction `ŷ` and true values `y`. + + | 0.5 * |ŷ - y|, for |ŷ - y| <= δ + Huber loss = | + | δ * (|ŷ - y| - 0.5 * δ), otherwise +""" +function huber_loss(ŷ, y; δ=eltype(ŷ)(1)) + abs_error = abs.(ŷ .- y) + temp = 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 + function _crossentropy(ŷ::AbstractVecOrMat, y::AbstractVecOrMat, weight::Nothing) return -sum(y .* log.(ŷ)) * 1 // size(y, 2) end @@ -17,22 +65,63 @@ function _crossentropy(ŷ::AbstractVecOrMat, y::AbstractVecOrMat, weight::Abstr return -sum(y .* log.(ŷ) .* weight) * 1 // size(y, 2) end +""" + crossentropy(ŷ, y; weight = nothing) + +Return the cross entropy between the given probability distributions; +calculated as `-sum(y .* log.(ŷ) .* weight) / size(y, 2)`. + +`weight` can be `Nothing`, a `Number` or an `AbstractVector`. +`weight=nothing` acts like `weight=1` but is faster. + +See also: [`Flux.logitcrossentropy`](@ref), [`Flux.binarycrossentropy`](@ref), [`Flux.logitbinarycrossentropy`](@ref) + +# Examples +```jldoctest +julia> Flux.crossentropy(softmax([-1.1491, 0.8619, 0.3127]), [1, 1, 0]) +3.085467254747739 +``` +""" crossentropy(ŷ::AbstractVecOrMat, y::AbstractVecOrMat; weight=nothing) = _crossentropy(ŷ, y, weight) -function logitcrossentropy(logŷ::AbstractVecOrMat, y::AbstractVecOrMat; weight = 1) - return -sum(y .* logsoftmax(logŷ) .* weight) * 1 // size(y, 2) +""" + logitcrossentropy(ŷ, y; weight = 1) + +Return the crossentropy computed after a [`Flux.logsoftmax`](@ref) operation; +calculated as `-sum(y .* logsoftmax(ŷ) .* weight) / size(y, 2)`. + +`logitcrossentropy(ŷ, y)` is mathematically equivalent to +[`Flux.crossentropy(softmax(log(ŷ)), y)`](@ref) but it is more numerically stable. + +See also: [`Flux.crossentropy`](@ref), [`Flux.binarycrossentropy`](@ref), [`Flux.logitbinarycrossentropy`](@ref) + +# Examples +```jldoctest +julia> Flux.logitcrossentropy([-1.1491, 0.8619, 0.3127], [1, 1, 0]) +3.085467254747738 +``` +""" +function logitcrossentropy(ŷ::AbstractVecOrMat, y::AbstractVecOrMat; weight = 1) + return -sum(y .* logsoftmax(ŷ) .* weight) * 1 // size(y, 2) end """ binarycrossentropy(ŷ, y; ϵ=eps(ŷ)) -Return `-y*log(ŷ + ϵ) - (1-y)*log(1-ŷ + ϵ)`. The ϵ term provides numerical stability. +Return ``-y*\\log(ŷ + ϵ) - (1-y)*\\log(1-ŷ + ϵ)``. The `ϵ` term provides numerical stability. - julia> binarycrossentropy.(σ.([-1.1491, 0.8619, 0.3127]), [1, 1, 0.]) - 3-element Array{Float64,1}: - 1.4244 - 0.352317 - 0.86167 +Typically, the prediction `ŷ` is given by the output of a [`sigmoid`](@ref) activation. + +See also: [`Flux.crossentropy`](@ref), [`Flux.logitcrossentropy`](@ref), [`Flux.logitbinarycrossentropy`](@ref) + +# Examples +```jldoctest +julia> Flux.binarycrossentropy.(σ.([-1.1491, 0.8619, 0.3127]), [1, 1, 0]) +3-element Array{Float64,1}: + 1.424397097347566 + 0.35231664672364077 + 0.8616703662235441 +``` """ binarycrossentropy(ŷ, y; ϵ=eps(ŷ)) = -y*log(ŷ + ϵ) - (1 - y)*log(1 - ŷ + ϵ) @@ -40,44 +129,52 @@ binarycrossentropy(ŷ, y; ϵ=eps(ŷ)) = -y*log(ŷ + ϵ) - (1 - y)*log(1 - ŷ CuArrays.@cufunc binarycrossentropy(ŷ, y; ϵ=eps(ŷ)) = -y*log(ŷ + ϵ) - (1 - y)*log(1 - ŷ + ϵ) """ - logitbinarycrossentropy(logŷ, y) + logitbinarycrossentropy(ŷ, y) -`logitbinarycrossentropy(logŷ, y)` is mathematically equivalent to `binarycrossentropy(σ(logŷ), y)` -but it is more numerically stable. +`logitbinarycrossentropy(ŷ, y)` is mathematically equivalent to +[`Flux.binarycrossentropy(σ(log(ŷ)), y)`](@ref) but it is more numerically stable. - julia> logitbinarycrossentropy.([-1.1491, 0.8619, 0.3127], [1, 1, 0.]) - 3-element Array{Float64,1}: - 1.4244 - 0.352317 - 0.86167 +See also: [`Flux.crossentropy`](@ref), [`Flux.logitcrossentropy`](@ref), [`Flux.binarycrossentropy`](@ref) + +# Examples +```jldoctest +julia> Flux.logitbinarycrossentropy.([-1.1491, 0.8619, 0.3127], [1, 1, 0]) +3-element Array{Float64,1}: + 1.4243970973475661 + 0.35231664672364094 + 0.8616703662235443 +``` """ -logitbinarycrossentropy(logŷ, y) = (1 - y)*logŷ - logσ(logŷ) +logitbinarycrossentropy(ŷ, y) = (1 - y)*ŷ - logσ(ŷ) # Re-definition to fix interaction with CuArrays. -CuArrays.@cufunc logitbinarycrossentropy(logŷ, y) = (1 - y)*logŷ - logσ(logŷ) +CuArrays.@cufunc logitbinarycrossentropy(ŷ, y) = (1 - y)*ŷ - logσ(ŷ) """ - normalise(x::AbstractArray; dims=1) + normalise(x; dims=1) -Normalises `x` to mean 0 and standard deviation 1, across the dimensions given by `dims`. Defaults to normalising over columns. +Normalise `x` to mean 0 and standard deviation 1 across the dimensions given by `dims`. +Defaults to normalising over columns. - julia> a = reshape(collect(1:9), 3, 3) - 3×3 Array{Int64,2}: - 1 4 7 - 2 5 8 - 3 6 9 +```jldoctest +julia> a = reshape(collect(1:9), 3, 3) +3×3 Array{Int64,2}: + 1 4 7 + 2 5 8 + 3 6 9 - julia> normalise(a) - 3×3 Array{Float64,2}: - -1.22474 -1.22474 -1.22474 - 0.0 0.0 0.0 - 1.22474 1.22474 1.22474 +julia> Flux.normalise(a) +3×3 Array{Float64,2}: + -1.22474 -1.22474 -1.22474 + 0.0 0.0 0.0 + 1.22474 1.22474 1.22474 - julia> normalise(a, dims=2) - 3×3 Array{Float64,2}: - -1.22474 0.0 1.22474 - -1.22474 0.0 1.22474 - -1.22474 0.0 1.22474 +julia> Flux.normalise(a, dims=2) +3×3 Array{Float64,2}: + -1.22474 0.0 1.22474 + -1.22474 0.0 1.22474 + -1.22474 0.0 1.22474 +``` """ function normalise(x::AbstractArray; dims=1) μ′ = mean(x, dims = dims) @@ -87,26 +184,81 @@ end """ kldivergence(ŷ, y) -KLDivergence is a measure of how much one probability distribution is different from the other. -It is always non-negative and zero only when both the distributions are equal everywhere. -[KL Divergence](https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence). + +Return the +[Kullback-Leibler divergence](https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence) +between the given probability distributions. + +KL divergence is a measure of how much one probability distribution is different +from the other. +It is always non-negative and zero only when both the distributions are equal +everywhere. """ function kldivergence(ŷ, y) - entropy = sum(y .* log.(y)) *1 //size(y,2) + entropy = sum(y .* log.(y)) * 1 //size(y,2) cross_entropy = crossentropy(ŷ, y) return entropy + cross_entropy end """ poisson(ŷ, y) -Poisson loss function is a measure of how the predicted distribution diverges from the expected distribution. -[Poisson Loss](https://peltarion.com/knowledge-center/documentation/modeling-view/build-an-ai-model/loss-functions/poisson). + +Return how much the predicted distribution `ŷ` diverges from the expected Poisson +distribution `y`; calculated as `sum(ŷ .- y .* log.(ŷ)) / size(y, 2)`. + +[More information.](https://peltarion.com/knowledge-center/documentation/modeling-view/build-an-ai-model/loss-functions/poisson). """ -poisson(ŷ, y) = sum(ŷ .- y .* log.(ŷ)) *1 // size(y,2) +poisson(ŷ, y) = sum(ŷ .- y .* log.(ŷ)) * 1 // size(y,2) """ hinge(ŷ, y) -Measures the loss given the prediction ŷ and true labels y(containing 1 or -1). -[Hinge Loss](https://en.wikipedia.org/wiki/Hinge_loss). + +Return the [hinge loss](https://en.wikipedia.org/wiki/Hinge_loss) given the +prediction `ŷ` and true labels `y` (containing 1 or -1); calculated as +`sum(max.(0, 1 .- ŷ .* y)) / size(y, 2)`. + +See also: [`squared_hinge`](@ref) """ -hinge(ŷ, y) = sum(max.(0, 1 .- ŷ .* y)) *1 // size(y,2) +hinge(ŷ, y) = sum(max.(0, 1 .- ŷ .* y)) * 1 // size(y, 2) + +""" + squared_hinge(ŷ, y) + +Return the squared hinge loss given the prediction `ŷ` and true labels `y` +(containing 1 or -1); calculated as `sum((max.(0, 1 .- ŷ .* y)).^2) / size(y, 2)`. + +See also: [`hinge`](@ref) +""" +squared_hinge(ŷ, y) = sum((max.(0, 1 .- ŷ .* y)).^2) * 1 // size(y, 2) + +""" + dice_coeff_loss(ŷ, y; smooth=1) + +Return a loss based on the dice coefficient. +Used in the [V-Net](https://arxiv.org/pdf/1606.04797v1.pdf) image segmentation +architecture. +Similar to the F1_score. Calculated as: + 1 - 2*sum(|ŷ .* y| + smooth) / (sum(ŷ.^2) + sum(y.^2) + smooth)` +""" +dice_coeff_loss(ŷ, y; smooth=eltype(ŷ)(1.0)) = 1 - (2*sum(y .* ŷ) + smooth) / (sum(y.^2) + sum(ŷ.^2) + smooth) + +""" + tversky_loss(ŷ, y; β=0.7) + +Return the [Tversky loss](https://arxiv.org/pdf/1706.05721.pdf). +Used with imbalanced data to give more weight to false negatives. +Larger β weigh recall higher than precision (by placing more emphasis on false negatives) +Calculated as: + 1 - sum(|y .* ŷ| + 1) / (sum(y .* ŷ + β*(1 .- y) .* ŷ + (1 - β)*y .* (1 .- ŷ)) + 1) +""" +tversky_loss(ŷ, y; β=eltype(ŷ)(0.7)) = 1 - (sum(y .* ŷ) + 1) / (sum(y .* ŷ + β*(1 .- y) .* ŷ + (1 - β)*y .* (1 .- ŷ)) + 1) + +""" + flatten(x::AbstractArray) + +Transform (w, h, c, b)-shaped input into (w × h × c, b)-shaped output +by linearizing all values for each element in the batch. +""" +function flatten(x::AbstractArray) + return reshape(x, :, size(x)[end]) +end diff --git a/src/onehot.jl b/src/onehot.jl index 7a3123ec..4b7e5e36 100644 --- a/src/onehot.jl +++ b/src/onehot.jl @@ -37,30 +37,28 @@ import Adapt: adapt, adapt_structure adapt_structure(T, xs::OneHotMatrix) = OneHotMatrix(xs.height, adapt(T, xs.data)) -import .CuArrays: CuArray, cudaconvert +import .CuArrays: CuArray, CuArrayStyle, cudaconvert import Base.Broadcast: BroadcastStyle, ArrayStyle -BroadcastStyle(::Type{<:OneHotMatrix{<:CuArray}}) = ArrayStyle{CuArray}() +BroadcastStyle(::Type{<:OneHotMatrix{<:CuArray}}) = CuArrayStyle{2}() cudaconvert(x::OneHotMatrix{<:CuArray}) = OneHotMatrix(x.height, cudaconvert(x.data)) """ onehot(l, labels[, unk]) -Create an [`OneHotVector`](@ref) wtih `l`-th element be `true` based on possible `labels` set. -If `unk` is given, it retruns `onehot(unk, labels)` if the input label `l` is not find in `labels`; otherwise -it will error. - -## Examples +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. +# Examples ```jldoctest -julia> using Flux: onehot - -julia> onehot(:b, [:a, :b, :c]) +julia> Flux.onehot(:b, [:a, :b, :c]) 3-element Flux.OneHotVector: 0 1 0 -julia> onehot(:c, [:a, :b, :c]) +julia> Flux.onehot(:c, [:a, :b, :c]) 3-element Flux.OneHotVector: 0 0 @@ -82,15 +80,14 @@ end """ onehotbatch(ls, labels[, unk...]) -Create an [`OneHotMatrix`](@ref) with a batch of labels based on possible `labels` set, returns the -`onehot(unk, labels)` if given labels `ls` is not found in set `labels`. - -## Examples +Create a `OneHotMatrix` with a batch of labels based on the +possible set of `labels`. +If `unk` is given, return [`onehot(unk, labels)`](@ref) if one of the input +labels `ls` is not found in `labels`; otherwise it will error. +# Examples ```jldoctest -julia> using Flux: onehotbatch - -julia> onehotbatch([:b, :a, :b], [:a, :b, :c]) +julia> Flux.onehotbatch([:b, :a, :b], [:a, :b, :c]) 3×3 Flux.OneHotMatrix{Array{Flux.OneHotVector,1}}: 0 1 0 1 0 1 @@ -107,13 +104,12 @@ Base.argmax(xs::OneHotVector) = xs.ix Inverse operations of [`onehot`](@ref). +# Examples ```jldoctest -julia> using Flux: onecold - -julia> onecold([true, false, false], [:a, :b, :c]) +julia> Flux.onecold([true, false, false], [:a, :b, :c]) :a -julia> onecold([0.3, 0.2, 0.5], [:a, :b, :c]) +julia> Flux.onecold([0.3, 0.2, 0.5], [:a, :b, :c]) :c ``` """ diff --git a/src/optimise/Optimise.jl b/src/optimise/Optimise.jl index 68c18a6f..5ffe4a8f 100644 --- a/src/optimise/Optimise.jl +++ b/src/optimise/Optimise.jl @@ -1,7 +1,7 @@ module Optimise -export train!, - SGD, Descent, ADAM, Momentum, Nesterov, RMSProp, +export train!, update!, + Descent, ADAM, Momentum, Nesterov, RMSProp, ADAGrad, AdaMax, ADADelta, AMSGrad, NADAM, ADAMW,RADAM, InvDecay, ExpDecay, WeightDecay, stop, Optimiser diff --git a/src/optimise/optimisers.jl b/src/optimise/optimisers.jl index cf4496f4..611edddb 100644 --- a/src/optimise/optimisers.jl +++ b/src/optimise/optimisers.jl @@ -6,24 +6,25 @@ const ϵ = 1e-8 # TODO: should use weak refs """ - Descent(η) + Descent(η = 0.1) Classic gradient descent optimiser with learning rate `η`. For each parameter `p` and its gradient `δp`, this runs `p -= η*δp` -## Parameters - - Learning Rate (η): The amount by which the gradients are discounted before updating the weights. Defaults to `0.1`. +# Parameters +- Learning rate (`η`): Amount by which gradients are discounted before updating + the weights. -## Example -```julia-repl -opt = Descent() # uses default η (0.1) +# Examples +```julia +opt = Descent() -opt = Descent(0.3) # use provided η +opt = Descent(0.3) ps = params(model) gs = gradient(ps) do - loss(x, y) + loss(x, y) end Flux.Optimise.update!(opt, ps, gs) @@ -40,17 +41,19 @@ function apply!(o::Descent, x, Δ) end """ - Momentum(η, ρ) + Momentum(η = 0.01, ρ = 0.9) -Gradient descent with learning rate `η` and momentum `ρ`. +Gradient descent optimizer with learning rate `η` and momentum `ρ`. -## Parameters - - Learning Rate (`η`): Amount by which gradients are discounted before updating the weights. Defaults to `0.01`. - - Momentum (`ρ`): Parameter that accelerates descent in the relevant direction and dampens oscillations. Defaults to `0.9`. +# Parameters +- Learning rate (`η`): Amount by which gradients are discounted before updating + the weights. +- Momentum (`ρ`): Controls the acceleration of gradient descent in the + prominent direction, in effect dampening oscillations. -## Examples +# Examples ```julia -opt = Momentum() # uses defaults of η = 0.01 and ρ = 0.9 +opt = Momentum() opt = Momentum(0.01, 0.99) ``` @@ -71,17 +74,19 @@ function apply!(o::Momentum, x, Δ) end """ - Nesterov(η, ρ) + Nesterov(η = 0.001, ρ = 0.9) -Gradient descent with learning rate `η` and Nesterov momentum `ρ`. +Gradient descent optimizer with learning rate `η` and Nesterov momentum `ρ`. -## Parameters - - Learning Rate (η): Amount by which the gradients are dicsounted berfore updating the weights. Defaults to `0.001`. - - Nesterov Momentum (ρ): Paramters controlling the amount of nesterov momentum to be applied. Defaults to `0.9`. +# Parameters +- Learning rate (`η`): Amount by which gradients are discounted before updating + the weights. +- Nesterov momentum (`ρ`): Controls the acceleration of gradient descent in the + prominent direction, in effect dampening oscillations. -## Examples +# Examples ```julia -opt = Nesterov() # uses defaults η = 0.001 and ρ = 0.9 +opt = Nesterov() opt = Nesterov(0.003, 0.95) ``` @@ -103,23 +108,25 @@ function apply!(o::Nesterov, x, Δ) end """ - RMSProp(η, ρ) + RMSProp(η = 0.001, ρ = 0.9) -Implements the RMSProp algortihm. Often a good choice for recurrent networks. Paramters other than learning rate generally don't need tuning. +Optimizer using the +[RMSProp](https://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf) +algorithm. Often a good choice for recurrent networks. Parameters other than learning rate +generally don't need tuning. -## Parameters - - Learning Rate (η): Defaults to `0.001`. - - Rho (ρ): Defaults to `0.9`. +# Parameters +- Learning rate (`η`): Amount by which gradients are discounted before updating + the weights. +- Momentum (`ρ`): Controls the acceleration of gradient descent in the + prominent direction, in effect dampening oscillations. -## Examples +# Examples ```julia -opt = RMSProp() # uses default η = 0.001 and ρ = 0.9 +opt = RMSProp() opt = RMSProp(0.002, 0.95) ``` - -## References -[RMSProp](https://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf) """ mutable struct RMSProp eta::Float64 @@ -137,23 +144,22 @@ function apply!(o::RMSProp, x, Δ) end """ - ADAM(η, β::Tuple) + ADAM(η = 0.001, β::Tuple = (0.9, 0.999)) -Implements the ADAM optimiser. +[ADAM](https://arxiv.org/abs/1412.6980v8) optimiser. -## Paramters - - Learning Rate (`η`): Defaults to `0.001`. - - Beta (`β::Tuple`): The first element refers to β1 and the second to β2. Defaults to `(0.9, 0.999)`. - -## Examples +# Parameters +- Learning rate (`η`): Amount by which gradients are discounted before updating + the weights. +- Decay of momentums (`β::Tuple`): Exponential decay for the first (β1) and the + second (β2) momentum estimate. +# Examples ```julia -opt = ADAM() # uses the default η = 0.001 and β = (0.9, 0.999) +opt = ADAM() opt = ADAM(0.001, (0.9, 0.8)) ``` -## References -[ADAM](https://arxiv.org/abs/1412.6980v8) optimiser. """ mutable struct ADAM eta::Float64 @@ -174,24 +180,22 @@ function apply!(o::ADAM, x, Δ) end """ - RADAM(η, β::Tuple) + RADAM(η = 0.001, β::Tuple = (0.9, 0.999)) -Implements the rectified ADAM optimizer. +[Rectified ADAM](https://arxiv.org/pdf/1908.03265v1.pdf) optimizer. -## Parameters - - Learning Rate (η): Defaults to `0.001` - - Beta (β::Tuple): The first element refers to β1 and the second to β2. Defaults to `(0.9, 0.999)`. - -## Examples +# Parameters +- Learning rate (`η`): Amount by which gradients are discounted before updating + the weights. +- Decay of momentums (`β::Tuple`): Exponential decay for the first (β1) and the + second (β2) momentum estimate. +# Examples ```julia -opt = RADAM() # uses the default η = 0.001 and β = (0.9, 0.999) +opt = RADAM() opt = RADAM(0.001, (0.9, 0.8)) ``` - -## References -[RADAM](https://arxiv.org/pdf/1908.03265v1.pdf) optimiser (Rectified ADAM). """ mutable struct RADAM eta::Float64 @@ -219,22 +223,22 @@ function apply!(o::RADAM, x, Δ) end """ - AdaMax(η, β::Tuple) + AdaMax(η = 0.001, β::Tuple = (0.9, 0.999)) -Variant of ADAM based on ∞-norm. +[AdaMax](https://arxiv.org/abs/1412.6980v9) is a variant of ADAM based on the ∞-norm. -## Parameters - - Learning Rate (η): Defaults to `0.001` - - Beta (β::Tuple): The first element refers to β1 and the second to β2. Defaults to `(0.9, 0.999)`. +# Parameters +- Learning rate (`η`): Amount by which gradients are discounted before updating + the weights. +- Decay of momentums (`β::Tuple`): Exponential decay for the first (β1) and the + second (β2) momentum estimate. -## Examples +# Examples ```julia -opt = AdaMax() # uses default η and β +opt = AdaMax() opt = AdaMax(0.001, (0.9, 0.995)) ``` -## References -[AdaMax](https://arxiv.org/abs/1412.6980v9) optimiser. """ mutable struct AdaMax eta::Float64 @@ -255,23 +259,22 @@ function apply!(o::AdaMax, x, Δ) end """ - ADAGrad(η) + ADAGrad(η = 0.1) -Implements AdaGrad. It has parameter specific learning rates based on how frequently it is updated. +[ADAGrad](http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) optimizer. It has +parameter specific learning rates based on how frequently it is updated. +Parameters don't need tuning. -## Parameters - - Learning Rate (η): Defaults to `0.1` +# Parameters +- Learning rate (`η`): Amount by which gradients are discounted before updating + the weights. -## Examples +# Examples ```julia -opt = ADAGrad() # uses default η = 0.1 +opt = ADAGrad() opt = ADAGrad(0.001) ``` - -## References -[ADAGrad](http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf) optimiser. -Parameters don't need tuning. """ mutable struct ADAGrad eta::Float64 @@ -288,21 +291,21 @@ function apply!(o::ADAGrad, x, Δ) end """ - ADADelta(ρ) + ADADelta(ρ = 0.9) -Version of ADAGrad that adapts learning rate based on a window of past gradient updates. Parameters don't need tuning. +[ADADelta](https://arxiv.org/abs/1212.5701) is a version of ADAGrad adapting its learning +rate based on a window of past gradient updates. +Parameters don't need tuning. -## Parameters - - Rho (ρ): Factor by which gradient is decayed at each time step. Defaults to `0.9`. +# Parameters +- Rho (`ρ`): Factor by which the gradient is decayed at each time step. -## Examples +# Examples ```julia -opt = ADADelta() # uses default ρ = 0.9 +opt = ADADelta() + opt = ADADelta(0.89) ``` - -## References -[ADADelta](https://arxiv.org/abs/1212.5701) optimiser. """ mutable struct ADADelta rho::Float64 @@ -321,22 +324,23 @@ function apply!(o::ADADelta, x, Δ) end """ - AMSGrad(η, β::Tuple) + AMSGrad(η = 0.001, β::Tuple = (0.9, 0.999)) -Implements AMSGrad version of the ADAM optimiser. Parameters don't need tuning. +The [AMSGrad](https://openreview.net/forum?id=ryQu7f-RZ) version of the ADAM +optimiser. Parameters don't need tuning. -## Parameters - - Learning Rate (η): Defaults to `0.001`. - - Beta (β::Tuple): The first element refers to β1 and the second to β2. Defaults to `(0.9, 0.999)`. +# Parameters +- Learning rate (`η`): Amount by which gradients are discounted before updating + the weights. +- Decay of momentums (`β::Tuple`): Exponential decay for the first (β1) and the + second (β2) momentum estimate. -## Examples +# Examples ```julia -opt = AMSGrad() # uses default η and β +opt = AMSGrad() + opt = AMSGrad(0.001, (0.89, 0.995)) ``` - -## References -[AMSGrad](https://openreview.net/forum?id=ryQu7f-RZ) optimiser. """ mutable struct AMSGrad eta::Float64 @@ -356,22 +360,23 @@ function apply!(o::AMSGrad, x, Δ) end """ - NADAM(η, β::Tuple) + NADAM(η = 0.001, β::Tuple = (0.9, 0.999)) -Nesterov variant of ADAM. Parameters don't need tuning. +[NADAM](http://cs229.stanford.edu/proj2015/054_report.pdf) is a Nesterov variant of ADAM. +Parameters don't need tuning. -## Parameters - - Learning Rate (η): Defaults to `0.001`. - - Beta (β::Tuple): The first element refers to β1 and the second to β2. Defaults to `(0.9, 0.999)`. +# Parameters +- Learning rate (`η`): Amount by which gradients are discounted before updating + the weights. +- Decay of momentums (`β::Tuple`): Exponential decay for the first (β1) and the + second (β2) momentum estimate. -## Examples +# Examples ```julia -opt = NADAM() # uses default η and β +opt = NADAM() + opt = NADAM(0.002, (0.89, 0.995)) ``` - -## References -[NADAM](http://cs229.stanford.edu/proj2015/054_report.pdf) optimiser. """ mutable struct NADAM eta::Float64 @@ -392,23 +397,24 @@ function apply!(o::NADAM, x, Δ) end """ - ADAMW(η, β::Tuple, decay) + ADAMW(η = 0.001, β::Tuple = (0.9, 0.999), decay = 0) -Variant of ADAM defined by fixing weight decay regularization. +[ADAMW](https://arxiv.org/abs/1711.05101) is a variant of ADAM fixing (as in repairing) its +weight decay regularization. -## Parameters - - Learning Rate (η): Defaults to `0.001`. - - Beta (β::Tuple): The first element refers to β1 and the second to β2. Defaults to (0.9, 0.999). - - decay: Decay applied to weights during optimisation. Defaults to 0. +# Parameters +- Learning rate (`η`): Amount by which gradients are discounted before updating + the weights. +- Decay of momentums (`β::Tuple`): Exponential decay for the first (β1) and the + second (β2) momentum estimate. +- `decay`: Decay applied to weights during optimisation. -## Examples +# Examples ```julia -opt = ADAMW() # uses default η, β and decay +opt = ADAMW() + opt = ADAMW(0.001, (0.89, 0.995), 0.1) ``` - -## References -[ADAMW](https://arxiv.org/abs/1711.05101) """ ADAMW(η = 0.001, β = (0.9, 0.999), decay = 0) = Optimiser(ADAM(η, β), WeightDecay(decay)) @@ -441,17 +447,15 @@ function apply!(o::Optimiser, x, Δ) end """ - InvDecay(γ) + InvDecay(γ = 0.001) -Applies inverse time decay to an optimiser, i.e., the effective step size at iteration `n` is `eta / (1 + γ * n)` where `eta` is the initial step size. The wrapped optimiser's step size is not modified. -``` +Apply inverse time decay to an optimiser, so that the effective step size at +iteration `n` is `eta / (1 + γ * n)` where `eta` is the initial step size. +The wrapped optimiser's step size is not modified. -## Parameters - - gamma (γ): Defaults to `0.001` - -## Example +# Examples ```julia - Optimiser(InvDecay(..), Opt(..)) +Optimiser(InvDecay(..), Opt(..)) ``` """ mutable struct InvDecay @@ -470,22 +474,25 @@ function apply!(o::InvDecay, x, Δ) end """ - ExpDecay(eta, decay, decay_step, clip) + ExpDecay(η = 0.001, decay = 0.1, decay_step = 1000, clip = 1e-4) -Discount the learning rate `eta` by a multiplicative factor `decay` every `decay_step` till a minimum of `clip`. +Discount the learning rate `η` by the factor `decay` every `decay_step` steps till +a minimum of `clip`. -## Parameters - - Learning Rate (eta): Defaults to `0.001`. - - decay: Factor by which the learning rate is discounted. Defaults to `0.1`. - - decay_step: Schedules decay operations by setting number of steps between two decay operations. Defaults to `1000`. - - clip: Minimum value of learning rate. Defaults to `1e-4`. +# Parameters +- Learning rate (`η`): Amount by which gradients are discounted before updating + the weights. +- `decay`: Factor by which the learning rate is discounted. +- `decay_step`: Schedule decay operations by setting the number of steps between + two decay operations. +- `clip`: Minimum value of learning rate. -## Example +# Examples To apply exponential decay to an optimiser: ```julia - Optimiser(ExpDecay(..), Opt(..)) +Optimiser(ExpDecay(..), Opt(..)) - opt = Optimiser(ExpDecay(), ADAM()) +opt = Optimiser(ExpDecay(), ADAM()) ``` """ mutable struct ExpDecay @@ -509,12 +516,12 @@ function apply!(o::ExpDecay, x, Δ) end """ - WeightDecay(wd) + WeightDecay(wd = 0) -Decays the weight by `wd` +Decay weights by `wd`. -## Parameters - - weight decay (wd): 0 +# Parameters +- Weight decay (`wd`) """ mutable struct WeightDecay wd::Real diff --git a/src/optimise/train.jl b/src/optimise/train.jl index ae0f334c..ade8e862 100644 --- a/src/optimise/train.jl +++ b/src/optimise/train.jl @@ -1,11 +1,26 @@ using Juno import Zygote: Params, gradient + + +""" + update!(x, x̄) + +Update the array `x` according to `x .-= x̄`. +""" function update!(x::AbstractArray, x̄) - x .+= x̄ - return x + x .-= x̄ end +""" + update!(opt, p, g) + update!(opt, ps::Params, gs) + +Perform an update step of the parameters `ps` (or the single parameter `p`) +according to optimizer `opt` and the gradients `gs` (the gradient `g`). + +As a result, the parameters are mutated and the optimizer's internal state may change. +""" function update!(opt, x, x̄) x .-= apply!(opt, x, x̄) end @@ -28,11 +43,10 @@ struct StopException <: Exception end stop() Call `Flux.stop()` in a callback to indicate when a callback condition is met. -This would trigger the train loop to stop and exit. +This will trigger the train loop to stop and exit. +# Examples ```julia -# Example callback: - cb = function () accuracy() > 0.9 && Flux.stop() end @@ -45,18 +59,19 @@ end """ train!(loss, params, data, opt; cb) -For each datapoint `d` in `data` computes the gradient of `loss(d...)` through -backpropagation and calls the optimizer `opt`. +For each datapoint `d` in `data` compute the gradient of `loss(d...)` through +backpropagation and call the optimizer `opt`. -Takes a callback as keyword argument `cb`. For example, this will print "training" -every 10 seconds: +In case datapoints `d` are of numeric array type, assume no splatting is needed +and compute the gradient of `loss(d)`. -```julia -Flux.train!(loss, params, data, opt, - cb = throttle(() -> println("training"), 10)) -``` +A callback is given with the keyword argument `cb`. For example, this will print +"training" every 10 seconds (using [`Flux.throttle`](@ref)): -The callback can call `Flux.stop()` to interrupt the training loop. + train!(loss, params, data, opt, + cb = throttle(() -> println("training"), 10)) + +The callback can call [`Flux.stop`](@ref) to interrupt the training loop. Multiple optimisers and callbacks can be passed to `opt` and `cb` as arrays. """ @@ -65,8 +80,14 @@ function train!(loss, ps, data, opt; cb = () -> ()) cb = runall(cb) @progress for d in data try - gs = gradient(ps) do - loss(d...) + if d isa AbstractArray{<:Number} + gs = gradient(ps) do + loss(d) + end + else + gs = gradient(ps) do + loss(d...) + end end update!(opt, ps, gs) cb() @@ -86,11 +107,12 @@ end Run `body` `N` times. Mainly useful for quickly doing multiple epochs of training in a REPL. -```julia -julia> @epochs 2 println("hello") -INFO: Epoch 1 +# Examples +```jldoctest +julia> Flux.@epochs 2 println("hello") +[ Info: Epoch 1 hello -INFO: Epoch 2 +[ Info: Epoch 2 hello ``` """ diff --git a/src/utils.jl b/src/utils.jl index 5e8eb270..c321bd91 100644 --- a/src/utils.jl +++ b/src/utils.jl @@ -1,10 +1,40 @@ # Arrays -nfan() = 1, 1 #fan_in, fan_out -nfan(n) = 1, n #A vector is treated as a n×1 matrix -nfan(n_out, n_in) = n_in, n_out #In case of Dense kernels: arranged as matrices -nfan(dims...) = prod(dims[1:end-2]) .* (dims[end-1], dims[end]) #In case of convolution kernels +nfan() = 1, 1 # fan_in, fan_out +nfan(n) = 1, n # A vector is treated as a n×1 matrix +nfan(n_out, n_in) = n_in, n_out # In case of Dense kernels: arranged as matrices +nfan(dims...) = prod(dims[1:end-2]) .* (dims[end-1], dims[end]) # In case of convolution kernels +""" + glorot_uniform(dims...) + +Return an `Array` of size `dims` containing random variables taken from a uniform +distribution in the interval ``[-x, x]``, where `x = sqrt(24 / sum(dims)) / 2`. + +# Examples +```jldoctest; setup = :(using Random; Random.seed!(0)) +julia> Flux.glorot_uniform(2, 3) +2×3 Array{Float32,2}: + 0.601094 -0.57414 -0.814925 + 0.900868 0.805994 0.057514 +``` +""" glorot_uniform(dims...) = (rand(Float32, dims...) .- 0.5f0) .* sqrt(24.0f0 / sum(nfan(dims...))) + +""" + glorot_normal(dims...) + +Return an `Array` of size `dims` containing random variables taken from a normal +distribution with mean 0 and standard deviation `sqrt(2 / sum(dims))`. + +# Examples +```jldoctest; setup = :(using Random; Random.seed!(0)) +julia> Flux.glorot_normal(3, 2) +3×2 Array{Float32,2}: + 0.429505 -0.0852891 + 0.523935 0.371009 + -0.223261 0.188052 +``` +""" glorot_normal(dims...) = randn(Float32, dims...) .* sqrt(2.0f0 / sum(nfan(dims...))) ones(T::Type, dims...) = Base.ones(T, dims...) @@ -13,9 +43,81 @@ zeros(T::Type, dims...) = Base.zeros(T, dims...) ones(dims...) = Base.ones(Float32, dims...) zeros(dims...) = Base.zeros(Float32, dims...) +""" + unsqueeze(xs, dim) + +Return `xs` reshaped into an `Array` one dimensionality higher than `xs`, +where `dim` indicates in which dimension `xs` is extended. + +# Examples +```jldoctest +julia> xs = [[1, 2], [3, 4], [5, 6]] +3-element Array{Array{Int64,1},1}: + [1, 2] + [3, 4] + [5, 6] + +julia> Flux.unsqueeze(xs, 1) +1×3 Array{Array{Int64,1},2}: + [1, 2] [3, 4] [5, 6] + +julia> Flux.unsqueeze([1 2; 3 4], 2) +2×1×2 Array{Int64,3}: +[:, :, 1] = + 1 + 3 + +[:, :, 2] = + 2 + 4 +``` +""" unsqueeze(xs, dim) = reshape(xs, (size(xs)[1:dim-1]..., 1, size(xs)[dim:end]...)) +""" + stack(xs, dim) + +Concatenate the given `Array` of `Array`s `xs` into a single `Array` along the +given dimension `dim`. + +# Examples +```jldoctest +julia> xs = [[1, 2], [3, 4], [5, 6]] +3-element Array{Array{Int64,1},1}: + [1, 2] + [3, 4] + [5, 6] + +julia> Flux.stack(xs, 1) +3×2 Array{Int64,2}: + 1 2 + 3 4 + 5 6 + +julia> cat(xs, dims=1) +3-element Array{Array{Int64,1},1}: + [1, 2] + [3, 4] + [5, 6] +``` +""" stack(xs, dim) = cat(unsqueeze.(xs, dim)..., dims=dim) + +""" + unstack(xs, dim) + +Unroll the given `xs` into an `Array` of `Array`s along the given dimension `dim`. + +# Examples +```jldoctest +julia> Flux.unstack([1 3 5 7; 2 4 6 8], 2) +4-element Array{Array{Int64,1},1}: + [1, 2] + [3, 4] + [5, 6] + [7, 8] +``` +""" unstack(xs, dim) = [copy(selectdim(xs, dim, i)) for i in 1:size(xs, dim)] """ @@ -23,9 +125,16 @@ unstack(xs, dim) = [copy(selectdim(xs, dim, i)) for i in 1:size(xs, dim)] Split `xs` into `n` parts. -```julia -julia> chunk(1:10, 3) -3-element Array{Array{Int64,1},1}: +# Examples +```jldoctest +julia> Flux.chunk(1:10, 3) +3-element Array{UnitRange{Int64},1}: + 1:4 + 5:8 + 9:10 + +julia> Flux.chunk(collect(1:10), 3) +3-element Array{SubArray{Int64,1,Array{Int64,1},Tuple{UnitRange{Int64}},true},1}: [1, 2, 3, 4] [5, 6, 7, 8] [9, 10] @@ -40,11 +149,12 @@ batchindex(xs, i) = (reverse(Base.tail(reverse(axes(xs))))..., i) Count the number of times that each element of `xs` appears. -```julia -julia> frequencies(['a','b','b']) +# Examples +```jldoctest +julia> Flux.frequencies(['a','b','b']) Dict{Char,Int64} with 2 entries: - 'b' => 2 'a' => 1 + 'b' => 2 ``` """ function frequencies(xs) @@ -60,12 +170,13 @@ head(x::Tuple) = reverse(Base.tail(reverse(x))) squeezebatch(x) = reshape(x, head(size(x))) """ - batch(xs) + batch(xs) Batch the arrays in `xs` into a single array. -```julia -julia> batch([[1,2,3],[4,5,6]]) +# Examples +```jldoctest +julia> Flux.batch([[1,2,3],[4,5,6]]) 3×2 Array{Int64,2}: 1 4 2 5 @@ -82,6 +193,25 @@ function batch(xs) return data end +""" +Return the given sequence padded with `p` up to a maximum length of `n`. + +# Examples +```jldoctest +julia> rpad([1, 2], 4, 0) +4-element Array{Int64,1}: + 1 + 2 + 0 + 0 + +julia> rpad([1, 2, 3], 2, 0) +3-element Array{Int64,1}: + 1 + 2 + 3 +``` +""" Base.rpad(v::AbstractVector, n::Integer, p) = [v; fill(p, max(n - length(v), 0))] """ @@ -90,8 +220,9 @@ Base.rpad(v::AbstractVector, n::Integer, p) = [v; fill(p, max(n - length(v), 0)) Take a list of `N` sequences, and turn them into a single sequence where each item is a batch of `N`. Short sequences will be padded by `pad`. -```julia -julia> batchseq([[1, 2, 3], [4, 5]], 0) +# Examples +```jldoctest +julia> Flux.batchseq([[1, 2, 3], [4, 5]], 0) 3-element Array{Array{Int64,1},1}: [1, 4] [2, 5] @@ -148,11 +279,15 @@ end # Other """ -Returns a function that when invoked, will only be triggered at most once -during `timeout` seconds. Normally, the throttled function will run -as much as it can, without ever going more than once per `wait` duration; -but if you'd like to disable the execution on the leading edge, pass -`leading=false`. To enable execution on the trailing edge, ditto. + throttle(f, timeout; leading=true, trailing=false) + +Return a function that when invoked, will only be triggered at most once +during `timeout` seconds. + +Normally, the throttled function will run as much as it can, without ever +going more than once per `wait` duration; but if you'd like to disable the +execution on the leading edge, pass `leading=false`. To enable execution on +the trailing edge, pass `trailing=true`. """ function throttle(f, timeout; leading=true, trailing=false) cooldown = true diff --git a/test/cuda/cuda.jl b/test/cuda/cuda.jl index c75cfb4e..128e5c7d 100644 --- a/test/cuda/cuda.jl +++ b/test/cuda/cuda.jl @@ -58,6 +58,13 @@ end @test y[3,:] isa CuArray end +@testset "restructure gpu" begin + dudt = Dense(1,1) |> gpu + p,re = Flux.destructure(dudt) + foo(x) = sum(re(p)(x)) + @test gradient(foo, cu(rand(1)))[1] isa CuArray +end + if CuArrays.has_cudnn() @info "Testing Flux/CUDNN" include("cudnn.jl") diff --git a/test/data.jl b/test/data.jl index 6b777873..20492323 100644 --- a/test/data.jl +++ b/test/data.jl @@ -1,22 +1,86 @@ -using Flux.Data -using Test +@testset "DataLoader" begin + X = reshape([1:10;], (2, 5)) + Y = [1:5;] -@test cmudict()["CATASTROPHE"] == :[K,AH0,T,AE1,S,T,R,AH0,F,IY0].args + d = DataLoader(X, batchsize=2) + batches = collect(d) + @test length(batches) == 3 + @test batches[1] == X[:,1:2] + @test batches[2] == X[:,3:4] + @test batches[3] == X[:,5:5] -@test length(CMUDict.phones()) == 39 + d = DataLoader(X, batchsize=2, partial=false) + batches = collect(d) + @test length(batches) == 2 + @test batches[1] == X[:,1:2] + @test batches[2] == X[:,3:4] -@test length(CMUDict.symbols()) == 84 + d = DataLoader(X, Y, batchsize=2) + batches = collect(d) + @test length(batches) == 3 + @test length(batches[1]) == 2 + @test length(batches[2]) == 2 + @test length(batches[3]) == 2 + @test batches[1][1] == X[:,1:2] + @test batches[1][2] == Y[1:2] + @test batches[2][1] == X[:,3:4] + @test batches[2][2] == Y[3:4] + @test batches[3][1] == X[:,5:5] + @test batches[3][2] == Y[5:5] -@test MNIST.images()[1] isa Matrix -@test MNIST.labels() isa Vector{Int64} + # test interaction with `train!` + θ = ones(2) + X = zeros(2, 10) + loss(x) = sum((x .- θ).^2) + d = DataLoader(X) + Flux.train!(loss, [θ], ncycle(d, 10), Descent(0.1)) + @test norm(θ) < 1e-4 -@test FashionMNIST.images()[1] isa Matrix -@test FashionMNIST.labels() isa Vector{Int64} + # test interaction with `train!` + θ = zeros(2) + X = ones(2, 10) + Y = fill(2, 10) + loss(x, y) = sum((y - x'*θ).^2) + d = DataLoader(X, Y) + Flux.train!(loss, [θ], ncycle(d, 10), Descent(0.1)) + @test norm(θ .- 1) < 1e-10 +end -@test Data.Sentiment.train() isa Vector{Data.Tree{Any}} +@testset "CMUDict" begin + @test cmudict()["CATASTROPHE"] == :[K,AH0,T,AE1,S,T,R,AH0,F,IY0].args -@test Iris.features() isa Matrix -@test size(Iris.features()) == (4,150) + @test length(CMUDict.phones()) == 39 -@test Iris.labels() isa Vector{String} -@test size(Iris.labels()) == (150,) + @test length(CMUDict.symbols()) == 84 +end + +@testset "MNIST" begin + @test MNIST.images()[1] isa Matrix + @test MNIST.labels() isa Vector{Int64} +end + +@testset "FashionMNIST" begin + @test FashionMNIST.images()[1] isa Matrix + @test FashionMNIST.labels() isa Vector{Int64} +end + +@testset "Sentiment" begin + @test Data.Sentiment.train() isa Vector{Data.Tree{Any}} +end + +@testset "Iris" begin + @test Iris.features() isa Matrix + @test size(Iris.features()) == (4,150) + + @test Iris.labels() isa Vector{String} + @test size(Iris.labels()) == (150,) +end + + +@testset "Housing" begin + @test Housing.features() isa Matrix # test broken due to SSL certifate expiration problem + @test size(Housing.features()) == (506, 13) + + @test Housing.targets() isa Array{Float64} + @test size(Housing.targets()) == (506, 1) +end diff --git a/test/layers/conv.jl b/test/layers/conv.jl index 1ba92c9f..8c825bfd 100644 --- a/test/layers/conv.jl +++ b/test/layers/conv.jl @@ -4,6 +4,10 @@ using Flux: gradient @testset "Pooling" begin x = randn(Float32, 10, 10, 3, 2) + gmp = GlobalMaxPool() + @test size(gmp(x)) == (1, 1, 3, 2) + gmp = GlobalMeanPool() + @test size(gmp(x)) == (1, 1, 3, 2) mp = MaxPool((2, 2)) @test mp(x) == maxpool(x, PoolDims(x, 2)) mp = MeanPool((2, 2)) @@ -187,4 +191,28 @@ end @test Flux.outdims(m, (5, 5)) == (4, 4) m = MeanPool((2, 2); stride = 2, pad = 3) @test Flux.outdims(m, (5, 5)) == (5, 5) -end \ No newline at end of file +end + +@testset "$ltype SamePad kernelsize $k" for ltype in (Conv, ConvTranspose, DepthwiseConv, CrossCor), k in ( (1,), (2,), (3,), (4,5), (6,7,8)) + data = ones(Float32, (k .+ 3)..., 1,1) + l = ltype(k, 1=>1, pad=SamePad()) + @test size(l(data)) == size(data) + + l = ltype(k, 1=>1, pad=SamePad(), dilation = k .÷ 2) + @test size(l(data)) == size(data) + + stride = 3 + l = ltype(k, 1=>1, pad=SamePad(), stride = stride) + if ltype == ConvTranspose + @test size(l(data))[1:end-2] == stride .* size(data)[1:end-2] .- stride .+ 1 + else + @test size(l(data))[1:end-2] == ceil.(Int, size(data)[1:end-2] ./ stride) + end +end + +@testset "$ltype SamePad windowsize $k" for ltype in (MeanPool, MaxPool), k in ( (1,), (2,), (3,), (4,5), (6,7,8)) + data = ones(Float32, (k .+ 3)..., 1,1) + + l = ltype(k, pad=SamePad()) + @test size(l(data))[1:end-2] == ceil.(Int, size(data)[1:end-2] ./ k) +end diff --git a/test/layers/normalisation.jl b/test/layers/normalisation.jl index 4399a256..ed2879b0 100644 --- a/test/layers/normalisation.jl +++ b/test/layers/normalisation.jl @@ -1,30 +1,32 @@ using Flux, Test, Statistics using Zygote: pullback -trainmode(f, x...) = pullback(f, x...)[1] -trainmode(f) = (x...) -> trainmode(f, x...) +evalwgrad(f, x...) = pullback(f, x...)[1] @testset "Dropout" begin x = [1.,2.,3.] @test x == Dropout(0.1)(x) - @test x == trainmode(Dropout(0), x) - @test zero(x) == trainmode(Dropout(1), x) + @test x == evalwgrad(Dropout(0), x) + @test zero(x) == evalwgrad(Dropout(1), x) x = rand(100) m = Dropout(0.9) - y = trainmode(m, x) + y = evalwgrad(m, x) @test count(a->a==0, y) > 50 - y = m(x) + testmode!(m, true) + y = evalwgrad(m, x) # should override istraining @test count(a->a==0, y) == 0 - y = trainmode(m, x) + testmode!(m, false) + y = evalwgrad(m, x) @test count(a->a==0, y) > 50 x = rand(Float32, 100) m = Chain(Dense(100,100), Dropout(0.9)) - y = trainmode(m, x) + y = evalwgrad(m, x) @test count(a->a == 0, y) > 50 - y = m(x) + testmode!(m, true) + y = evalwgrad(m, x) # should override istraining @test count(a->a == 0, y) == 0 x = rand(100, 50) @@ -49,7 +51,7 @@ end # initial m.σ is 1 # initial m.μ is 0 - y = trainmode(m, x) + y = evalwgrad(m, x) @test isapprox(y, [-1.22474 0 1.22474; -1.22474 0 1.22474], atol = 1.0e-5) # julia> x # 2×3 Array{Float64,2}: @@ -82,19 +84,19 @@ end @test isapprox(y, sigmoid.((x .- m.μ) ./ sqrt.(m.σ² .+ m.ϵ)), atol = 1.0e-7) end - let m = trainmode(BatchNorm(2)), x = reshape(Float32.(1:6), 3, 2, 1) + let m = trainmode!(BatchNorm(2)), x = reshape(Float32.(1:6), 3, 2, 1) y = reshape(permutedims(x, [2, 1, 3]), 2, :) y = permutedims(reshape(m(y), 2, 3, 1), [2, 1, 3]) @test m(x) == y end - let m = trainmode(BatchNorm(2)), x = reshape(Float32.(1:12), 2, 3, 2, 1) + let m = trainmode!(BatchNorm(2)), x = reshape(Float32.(1:12), 2, 3, 2, 1) y = reshape(permutedims(x, [3, 1, 2, 4]), 2, :) y = permutedims(reshape(m(y), 2, 2, 3, 1), [2, 3, 1, 4]) @test m(x) == y end - let m = trainmode(BatchNorm(2)), x = reshape(Float32.(1:24), 2, 2, 3, 2, 1) + let m = trainmode!(BatchNorm(2)), x = reshape(Float32.(1:24), 2, 2, 3, 2, 1) y = reshape(permutedims(x, [4, 1, 2, 3, 5]), 2, :) y = permutedims(reshape(m(y), 2, 2, 2, 3, 1), [2, 3, 4, 1, 5]) @test m(x) == y @@ -117,7 +119,7 @@ end x = Float64.(x) @test m.β == [0, 0] # initβ(2) @test m.γ == [1, 1] # initγ(2) - y = trainmode(m, x) + y = evalwgrad(m, x) #julia> x #[:, :, 1] = @@ -162,7 +164,7 @@ end @test isapprox(y, sigmoid.((x .- expand_inst(m.μ, affine_shape)) ./ sqrt.(expand_inst(m.σ², affine_shape) .+ m.ϵ)), atol = 1.0e-7) end - let m = trainmode(InstanceNorm(2)), sizes = (2, 4, 1, 2, 3), + let m = trainmode!(InstanceNorm(2)), sizes = (2, 4, 1, 2, 3), x = Float32.(reshape(collect(1:prod(sizes)), sizes)) y = reshape(permutedims(x, [3, 1, 2, 4, 5]), :, 2, 3) y = reshape(m(y), sizes...) @@ -172,14 +174,14 @@ end # check that μ, σ², and the output are the correct size for higher rank tensors let m = InstanceNorm(2), sizes = (5, 5, 3, 4, 2, 6), x = reshape(Float32.(collect(1:prod(sizes))), sizes) - y = trainmode(m, x) + y = evalwgrad(m, x) @test size(m.μ) == (sizes[end - 1], ) @test size(m.σ²) == (sizes[end - 1], ) @test size(y) == sizes end # show that instance norm is equal to batch norm when channel and batch dims are squashed - let m_inorm = trainmode(InstanceNorm(2)), m_bnorm = trainmode(BatchNorm(12)), sizes = (5, 5, 3, 4, 2, 6), + let m_inorm = trainmode!(InstanceNorm(2)), m_bnorm = trainmode!(BatchNorm(12)), sizes = (5, 5, 3, 4, 2, 6), x = reshape(Float32.(collect(1:prod(sizes))), sizes) @test m_inorm(x) == reshape(m_bnorm(reshape(x, (sizes[1:end - 2]..., :, 1))), sizes) end @@ -204,7 +206,7 @@ if VERSION >= v"1.1" @test m.β == [0, 0, 0, 0] # initβ(32) @test m.γ == [1, 1, 1, 1] # initγ(32) - y = trainmode(m, x) + y = evalwgrad(m, x) #julia> x #[:, :, 1] = @@ -263,7 +265,7 @@ if VERSION >= v"1.1" @test isapprox(y, out, atol = 1.0e-7) end - let m = trainmode(GroupNorm(2,2)), sizes = (2, 4, 1, 2, 3), + let m = trainmode!(GroupNorm(2,2)), sizes = (2, 4, 1, 2, 3), x = Float32.(reshape(collect(1:prod(sizes)), sizes)) y = reshape(permutedims(x, [3, 1, 2, 4, 5]), :, 2, 3) y = reshape(m(y), sizes...) @@ -273,20 +275,20 @@ if VERSION >= v"1.1" # check that μ, σ², and the output are the correct size for higher rank tensors let m = GroupNorm(4,2), sizes = (5, 5, 3, 4, 4, 6), x = Float32.(reshape(collect(1:prod(sizes)), sizes)) - y = trainmode(m, x) + y = evalwgrad(m, x) @test size(m.μ) == (m.G,1) @test size(m.σ²) == (m.G,1) @test size(y) == sizes end # show that group norm is the same as instance norm when the group size is the same as the number of channels - let IN = trainmode(InstanceNorm(4)), GN = trainmode(GroupNorm(4,4)), sizes = (2,2,3,4,5), + let IN = trainmode!(InstanceNorm(4)), GN = trainmode!(GroupNorm(4,4)), sizes = (2,2,3,4,5), x = Float32.(reshape(collect(1:prod(sizes)), sizes)) @test IN(x) ≈ GN(x) end # show that group norm is the same as batch norm for a group of size 1 and batch of size 1 - let BN = trainmode(BatchNorm(4)), GN = trainmode(GroupNorm(4,4)), sizes = (2,2,3,4,1), + let BN = trainmode!(BatchNorm(4)), GN = trainmode!(GroupNorm(4,4)), sizes = (2,2,3,4,1), x = Float32.(reshape(collect(1:prod(sizes)), sizes)) @test BN(x) ≈ GN(x) end diff --git a/test/layers/stateless.jl b/test/layers/stateless.jl index 7cb8ed2e..ebcd815c 100644 --- a/test/layers/stateless.jl +++ b/test/layers/stateless.jl @@ -1,6 +1,6 @@ using Test using Flux: onehotbatch, mse, crossentropy, logitcrossentropy, - σ, binarycrossentropy, logitbinarycrossentropy + σ, binarycrossentropy, logitbinarycrossentropy, flatten const ϵ = 1e-7 @@ -12,6 +12,20 @@ const ϵ = 1e-7 @testset "mse" begin @test mse(ŷ, y) ≈ (.1^2 + .9^2)/2 end + + @testset "mae" begin + @test Flux.mae(ŷ, y) ≈ 1/2 + end + + @testset "huber_loss" begin + @test Flux.huber_loss(ŷ, y) ≈ 0.20500000000000002 + end + + y = [123.0,456.0,789.0] + ŷ = [345.0,332.0,789.0] + @testset "msle" begin + @test Flux.msle(ŷ, y) ≈ 0.38813985859136585 + end # Now onehot y's y = onehotbatch([1, 1, 0, 0], 0:1) @@ -51,31 +65,50 @@ const ϵ = 1e-7 end y = [1 2 3] - y1 = [4.0 5.0 6.0] + ŷ = [4.0 5.0 6.0] @testset "kldivergence" begin - @test Flux.kldivergence(y, y1) ≈ 4.761838062403337 + @test Flux.kldivergence(ŷ, y) ≈ -1.7661057888493457 @test Flux.kldivergence(y, y) ≈ 0 end y = [1 2 3 4] - y1 = [5.0 6.0 7.0 8.0] + ŷ = [5.0 6.0 7.0 8.0] @testset "hinge" begin - @test Flux.hinge(y, y1) ≈ 0 + @test Flux.hinge(ŷ, y) ≈ 0 @test Flux.hinge(y, 0.5 .* y) ≈ 0.125 end + @testset "squared_hinge" begin + @test Flux.squared_hinge(ŷ, y) ≈ 0 + @test Flux.squared_hinge(y, 0.5 .* y) ≈ 0.0625 + end + y = [0.1 0.2 0.3] - y1 = [0.4 0.5 0.6] + ŷ = [0.4 0.5 0.6] @testset "poisson" begin - @test Flux.poisson(y, y1) ≈ 1.0160455586700767 + @test Flux.poisson(ŷ, y) ≈ 0.6278353988097339 @test Flux.poisson(y, y) ≈ 0.5044459776946685 end + y = [1.0 0.5 0.3 2.4] + ŷ = [0 1.4 0.5 1.2] + @testset "dice_coeff_loss" begin + @test Flux.dice_coeff_loss(ŷ, y) ≈ 0.2799999999999999 + @test Flux.dice_coeff_loss(y, y) ≈ 0.0 + end + + @testset "tversky_loss" begin + @test Flux.tversky_loss(ŷ, y) ≈ -0.06772009029345383 + @test Flux.tversky_loss(ŷ, y, β = 0.8) ≈ -0.09490740740740744 + @test Flux.tversky_loss(y, y) ≈ -0.5576923076923075 + end + @testset "no spurious promotions" begin for T in (Float32, Float64) y = rand(T, 2) ŷ = rand(T, 2) - for f in (mse, crossentropy, logitcrossentropy, Flux.kldivergence, Flux.hinge, Flux.poisson) + for f in (mse, crossentropy, logitcrossentropy, Flux.kldivergence, Flux.hinge, Flux.poisson, + Flux.mae, Flux.huber_loss, Flux.msle, Flux.squared_hinge, Flux.dice_coeff_loss, Flux.tversky_loss) fwd, back = Flux.pullback(f, ŷ, y) @test fwd isa T @test eltype(back(one(T))[1]) == T @@ -83,3 +116,10 @@ const ϵ = 1e-7 end end end + +@testset "helpers" begin + @testset "flatten" begin + x = randn(Float32, 10, 10, 3, 2) + @test size(flatten(x)) == (300, 2) + end +end diff --git a/test/runtests.jl b/test/runtests.jl index 1505e96a..c2ea0715 100644 --- a/test/runtests.jl +++ b/test/runtests.jl @@ -1,32 +1,50 @@ -using Flux, Test, Random, Statistics, Documenter -using Random +using Flux +using Flux.Data +using Test +using Random, Statistics, LinearAlgebra +using Documenter +using IterTools: ncycle Random.seed!(0) @testset "Flux" begin -@info "Testing Basics" + @testset "Utils" begin + include("utils.jl") + end -include("utils.jl") -include("onehot.jl") -include("optimise.jl") -include("data.jl") + @testset "Onehot" begin + include("onehot.jl") + end -@info "Testing Layers" + @testset "Optimise" begin + include("optimise.jl") + end -include("layers/basic.jl") -include("layers/normalisation.jl") -include("layers/stateless.jl") -include("layers/conv.jl") + @testset "Data" begin + include("data.jl") + end -if Flux.use_cuda[] - include("cuda/cuda.jl") -else - @warn "CUDA unavailable, not testing GPU support" -end + @testset "Layers" begin + include("layers/basic.jl") + include("layers/normalisation.jl") + include("layers/stateless.jl") + include("layers/conv.jl") + end -if VERSION >= v"1.2" - doctest(Flux) -end + @testset "CUDA" begin + if Flux.use_cuda[] + include("cuda/cuda.jl") + else + @warn "CUDA unavailable, not testing GPU support" + end + end -end + @testset "Docs" begin + if VERSION >= v"1.4" + DocMeta.setdocmeta!(Flux, :DocTestSetup, :(using Flux); recursive=true) + doctest(Flux) + end + end + +end # testset Flux