FasterAI (#151)

* Add dataset recipes and dataset registry

* Add fastai dataset registry with some recipes

* move `typify` helepr

* add learning method registry

* add test for `ImageSegmentationFolders` recipe

* add missing `mockblock` for `OneHotMulti`

* update learning methods

* add convenience `plotpredictions` method

* Add multi-label recipe

* add some tests for query functions

CHANGELOG.md

@@ -4,10 +4,12 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-## [Unreleased] - 2020-03-08 – now
+## [Unreleased] - 2020-07-32 – now
 
 ### Added
-- `plotlrfind` to visualize results of `LRFinderPhase`
+- High-level API "FasterAI"
+    - Find datasets and learning methods based on `Block`s: `finddataset`, `findlearningmethods`
+    - `loaddataset` for quickly loading data containers from configured recipes
+- Data container recipes (`DatasetRecipe`, `loadrecipe`)
 
-### Changed
-- Documentation notebooks to reflect changes in API
+### Changed

src/datablock/wrappers.jl

@@ -34,10 +34,10 @@ end
 Wrapper `Block` to attach a name to a block. Can be used in conjunction
 with [`Only`](#) to apply encodings to specific blocks only.
 """
-struct Named{Name, B<:Block} <: WrapperBlock
+struct Named{Name, B<:AbstractBlock} <: WrapperBlock
     block::B
 end
-Named(name::Symbol, block::B) where {B<:Block} = Named{name, B}(block)
+Named(name::Symbol, block::B) where {B<:AbstractBlock} = Named{name, B}(block)
 
 
 # the name is preserved through encodings and decodings

src/datasets/fastairegistry.jl

@@ -1,14 +1,15 @@
 
 const FASTAI_DATA_RECIPES = Dict{String, Vector{DatasetRecipe}}(
     # Image classification datasets
-    [name => [ImageClassificationFolders()] for name in (
+    [name => [ImageFolders()] for name in (
         "imagenette", "imagenette-160", "imagenette-320",
         "imagenette2", "imagenette2-160", "imagenette2-320",
         "imagewoof", "imagewoof-160", "imagewoof-320",
         "imagewoof2", "imagewoof2-160", "imagewoof2-320",
     )]...,
 
     "camvid_tiny" => [ImageSegmentationFolders()],
+    "pascal_2007" => [ImageTableMultiLabel()],
 )
 
 

src/datasets/recipes.jl

@@ -47,7 +47,7 @@ recipeblocks(::R) where {R<:DatasetRecipe} = recipeblocks(R)
 # ImageClfFolders
 
 """
-    ImageClfFolders(; labelfn = parentname, split = false)
+    ImageFolders(; labelfn = parentname, split = false)
 
 Recipe for loading a single-label image classification dataset
 stored in a hierarchical folder format. If `split == true`, split
@@ -56,16 +56,16 @@ defaults to the name of the parent folder but a custom function can
 be passed as `labelfn`.
 
 ```julia
-julia> recipeblocks(ImageClassificationFolders)
+julia> recipeblocks(ImageFolders)
 Tuple{Image{2}, Label}
 ```
 """
-Base.@kwdef struct ImageClassificationFolders <: DatasetRecipe
+Base.@kwdef struct ImageFolders <: DatasetRecipe
     labelfn = parentname
     split::Bool = false
 end
 
-function loadrecipe(recipe::ImageClassificationFolders, path)
+function loadrecipe(recipe::ImageFolders, path)
     isdir(path) || error("$path is not a directory")
     data = loadfolderdata(
         path,
@@ -81,7 +81,7 @@ function loadrecipe(recipe::ImageClassificationFolders, path)
     return data, blocks
 end
 
-recipeblocks(::Type{ImageClassificationFolders}) = Tuple{Image{2}, Label}
+recipeblocks(::Type{ImageFolders}) = Tuple{Image{2}, Label}
 
 
 # ImageSegmentationFolders
@@ -124,3 +124,37 @@ function loadrecipe(recipe::ImageSegmentationFolders, path)
 end
 
 recipeblocks(::Type{ImageSegmentationFolders}) = Tuple{Image{2}, Mask{2}}
+
+# ImageTableMultiLabel
+
+Base.@kwdef struct ImageTableMultiLabel <: DatasetRecipe
+    csvfile::String = "train.csv"
+    imagefolder::String = "train"
+    filecol::Symbol = :fname
+    labelcol::Symbol = :labels
+    split::Bool = false
+    splitcol::Symbol = :is_valid
+    labelsep::String = " "
+end
+
+
+function loadrecipe(recipe::ImageTableMultiLabel, path)
+    csvpath = joinpath(path, recipe.csvfile)
+    isfile(csvpath) || error("File $csvpath does not exist")
+    df = loadfile(csvpath)
+    images = mapobs(f -> loadfile(joinpath(path, recipe.imagefolder, f)), df[:, recipe.filecol])
+    labels = map(str -> split(str, recipe.labelsep), df[:,recipe.labelcol])
+    data = (images, labels)
+    blocks = Image{2}(), LabelMulti(unique(Iterators.flatten(labels)))
+    if recipe.split
+        idxs = 1:nobs(data)
+        splits = df[:, recipe.splitcol]
+        data = Dict(
+            "train" => datasubset(data, idxs[splits]),
+            "valid" => datasubset(data, idxs[(!).(splits)])
+        )
+    end
+    return data, blocks
+end
+
+recipeblocks(::Type{ImageTableMultiLabel}) = Tuple{Image{2}, LabelMulti}

src/datasets/transformations.jl

@@ -6,7 +6,7 @@ struct MappedData{F, D}
     data::D
 end
 
-Base.show(io::IO, data::MappedData) = print(io, "mapobs($(data.f), $(data.data))")
+Base.show(io::IO, data::MappedData) = print(io, "mapobs($(data.f), $(summary(data.data)))")
 Base.show(io::IO, data::MappedData{F, <:AbstractArray}) where F = print(io, "mapobs($(data.f), $(ShowLimit(data.data, limit=80)))")
 LearnBase.nobs(data::MappedData) = nobs(data.data)
 LearnBase.getobs(data::MappedData, idx::Int) = data.f(getobs(data.data, idx))

test/datablock.jl

@@ -1,15 +1,6 @@
-#include("imports.jl")
-using FastAI
-import FastAI: Block, Encoding, encode, decode, checkblock, encodedblock, decodedblock
-using FastAI: Label, LabelMulti, Mask, Image, ImageTensor, testencoding
-using FastAI: OneHot
-using Test
-using StaticArrays
-using Images
-using FastAI: grabbounds
-using Images
-
-##
+include("imports.jl")
+
+
 struct ABlock <: Block
     end
 checkblock(::ABlock, ::Int) = true
@@ -69,9 +60,7 @@ end
     testencoding(enc, block, image)
     @testset "randstate is shared" begin
         im1, im2 = encode(enc, Training(), (block, block), (image, image))
-        im3 = encode(enc, Training(), block, image)
         @test im1 ≈ im2
-        @test !(im1 == im3)
     end
 
     @testset "don't transform data that doesn't need to be resized" begin

test/datasets/recipes.jl

@@ -15,31 +15,31 @@ function testrecipe(recipe::Datasets.DatasetRecipe, data, blocks)
 end
 
 
-@testset ExtendedTestSet "ImageClassificationFolders" begin
+@testset ExtendedTestSet "ImageFolders" begin
     path = joinpath(datasetpath("mnist_var_size_tiny"), "train")
 
     @testset ExtendedTestSet "Basic configuration" begin
-        recipe = Datasets.ImageClassificationFolders()
+        recipe = Datasets.ImageFolders()
         data, blocks = loadrecipe(recipe, path)
         testrecipe(recipe, data, blocks)
         @test blocks[1] isa Image
         @test blocks[2].classes == ["3", "7"]
     end
 
     @testset ExtendedTestSet "Split configuration" begin
-        recipe = Datasets.ImageClassificationFolders(split=true)
+        recipe = Datasets.ImageFolders(split=true)
         data, blocks = loadrecipe(recipe, path)
         testrecipe(recipe, data["train"], blocks)
     end
 
     @testset ExtendedTestSet "Error cases" begin
         @testset ExtendedTestSet "Empty directory" begin
-            recipe = Datasets.ImageClassificationFolders(split=true)
+            recipe = Datasets.ImageFolders(split=true)
             @test_throws ErrorException loadrecipe(recipe, mktempdir())
         end
 
         @testset ExtendedTestSet "Only one label" begin
-            recipe = Datasets.ImageClassificationFolders(labelfn=x -> "1")
+            recipe = Datasets.ImageFolders(labelfn=x -> "1")
             @test_throws ErrorException loadrecipe(recipe, path)
         end
     end

test/datasets/registry.jl

@@ -23,7 +23,7 @@ include("../imports.jl")
 
     @testset ExtendedTestSet "registerrecipe!" begin
         @test_nowarn Datasets.registerrecipe!(
-            reg, "mnist_var_size_tiny", Datasets.ImageClassificationFolders())
+            reg, "mnist_var_size_tiny", Datasets.ImageFolders())
     end
 
     @testset ExtendedTestSet "finddatasets" begin

test/fasterai.jl

@@ -0,0 +1,14 @@
+include("imports.jl")
+
+
+
+@testset ExtendedTestSet "FasterAI" begin
+    @test length(listdatasources()) > 10
+
+    @test !isempty(finddatasets(blocks=(Image, Label)))
+    @test !isempty(finddatasets(blocks=(Image, LabelMulti)))
+    @test !isempty(finddatasets(blocks=(Image, Mask)))
+
+    @test ImageClassificationSingle ∈ findlearningmethods((Image, Label))
+    @test ImageClassificationMulti ∈ findlearningmethods((Image, LabelMulti))
+end

test/imports.jl

@@ -2,8 +2,8 @@
 using Colors: RGB, N0f8, Gray
 using FastAI
 using FastAI: ParamGroups, IndexGrouper, getgroup, DiscriminativeLRs, decay_optim
-using FastAI: Image, Keypoints, Mask, testencoding, Label, OneHot, ProjectiveTransforms,
-    encodedblock, decodedblock, encode, decode, mockblock
+import FastAI: Image, Keypoints, Mask, testencoding, Label, OneHot, ProjectiveTransforms,
+    encodedblock, decodedblock, encode, decode, mockblock, checkblock, Block, Encoding
 using FilePathsBase
 using FastAI.Datasets
 using DLPipelines

test/runtests.jl

@@ -4,6 +4,14 @@ include("imports.jl")
 ##
 
 @testset ExtendedTestSet "FastAI.jl" begin
+    @testset ExtendedTestSet "datablock.jl" begin
+        include("datablock.jl")
+    end
+
+    @testset ExtendedTestSet "fasterai.jl" begin
+        include("fasterai.jl")
+    end
+
     @testset ExtendedTestSet "encodings/" begin
         @testset ExtendedTestSet "projective.jl" begin
             include("encodings/projective.jl")