Merge pull request #89 from adrhill/ah/flip-hw

Add `PermuteDims` transformation

src/DataAugmentation.jl

@@ -56,6 +56,7 @@ export Item,
     Polygon,
     ToEltype,
     ImageToTensor,
+    PermuteDims,
     Normalize,
     NormalizeIntensity,
     MaskMulti,

src/preprocessing.jl

@@ -219,6 +219,49 @@ function _colorview(C::Type{<:Color}, img)
     return colorview(C, img)
 end
 
+# ### [`PermuteDims`](#)
+
+"""
+    PermuteDims(perm)
+
+Permute the dimensions of an `ArrayItem`.
+`perm` is a vector or a tuple specifying the permutation, whose length has to match the dimensionality of the `ArrayItem`s data.
+
+Refer to the `permutedims` documentation for examples of permutation vectors `perm`.
+
+Supports `apply!`.
+
+## Examples
+
+Preprocessing an image with 3 color channels.
+
+{cell=PermuteDims}
+```julia
+using DataAugmentation, Images
+image = Image(rand(RGB, 20, 20))
+
+# Turn image to tensor and permute dimensions 2 and 1
+# to convert HWC (height, width, channel) array to WHC (width, height, channel)
+tfms = ImageToTensor() |> PermuteDims(2, 1, 3)
+apply(tfms, image)
+```
+
+"""
+struct PermuteDims{N} <: Transform
+    perm::NTuple{N, Int}
+end
+PermuteDims(perm...) = PermuteDims(perm)
+
+function apply(tfm::PermuteDims, item::ArrayItem; randstate = nothing)
+    data = PermutedDimsArray(itemdata(item), tfm.perm)
+    return ArrayItem(data)
+end
+
+function apply!(buf, tfm::PermuteDims, item::ArrayItem; randstate = nothing)
+    permutedims!(itemdata(buf), itemdata(item), tfm.perm)
+    return buf
+end
+
 # OneHot encoding
 
 """

test/preprocessing.jl

@@ -62,6 +62,18 @@ end
     end
 end
 
+@testset ExtendedTestSet "PermuteDims" begin
+    tfm = PermuteDims(2, 1, 4, 3)
+    item1 = ArrayItem(rand(3, 4, 5, 6))
+    item2 = ArrayItem(rand(3, 4, 5, 6))
+    @test_nowarn apply(tfm, item1)
+    a = itemdata(apply(tfm, item1))
+    @test size(a) == (4, 3, 6, 5)
+
+    testapply(tfm, item1)
+    testapply!(tfm, item1, item2)
+end
+
 @testset ExtendedTestSet "OneHot" begin
     tfm = OneHot()
     mask = rand(1:4, 10, 10)