Either may not work properly with Pipeline

[johnnychen94]

For the following three seeming "equivalent" operations, the first one randomly changes the image size...

For img = testimage("cameraman"):

• augment(img, Either(Rotate90()) |> FlipY()) returns image of size (512, 512) or (513, 512)
• augment(img, Rotate90() |> FlipY()) returns image of size (512, 512)
• augment(augment(img, Either(Rotate90())), FlipY()) returns image of size (512, 512)

[johnnychen94]

Without any evidence, this issue might be related to the unrolling part of lazy operations in augment_impl.

A workaround is to insert a CacheImage operation after Either to "enforce eager mode", e.g., augment(img, Either(Rotate90()) |> CacheImage() |> FlipY())

[Evizero]

yea. Rotate90 and NoOp only share the affine option for lazy. and there is likely some rounding error when rotating

[barucden]

I could not reproduce on the current master. Is it possible that it has been fixed?

julia> using Augmentor, TestImages

julia> img = TestImages.testimage("cameraman");

julia> a(img) = augment(img, Either(Rotate90()) |> FlipY())
a (generic function with 1 method)

julia> b(img) = augment(img, Rotate90() |> FlipY())
b (generic function with 1 method)

julia> c(img) = augment(augment(img, Either(Rotate90())), FlipY())
c (generic function with 1 method)

julia> unique!([size(a(img)) for i in 1:1000])
1-element Vector{Tuple{Int64, Int64}}:
 (512, 512)

julia> unique!([size(b(img)) for i in 1:1000])
1-element Vector{Tuple{Int64, Int64}}:
 (512, 512)

julia> unique!([size(c(img)) for i in 1:1000])
1-element Vector{Tuple{Int64, Int64}}:
 (512, 512)

[johnnychen94]

Great! I didn't notice this was fixed but it might be related to the discretization trick in ImageTransformations: JuliaImages/ImageTransformations.jl#107 and JuliaImages/ImageTransformations.jl#114 This might only happen when the affine method is called. Another similar issue is #77

An upcoming change in ImageTransformations@v0.9 is JuliaImages/ImageTransformations.jl#143