random failure OffsetArray tests for Scale and Zoom

[johnnychen94]

This bug is identified during the upgrade of Augmentor #31 Since that PR is growing bigger and becoming hard to review, it's better to open an issue here and fix it separately.

julia> using Augmentor, ImageCore, OffsetArrays, ImageDistances

julia> img = rand(Gray{N0f8}, 64, 64);

julia> out1 = Augmentor.applyeager(Scale(1.5), img);

julia> out2 = Augmentor.applyeager(Scale(1.5), OffsetArray(img, -1, -1));

julia> out1 == out2
false

julia> euclidean(out1, out2)
0.044884406f0

This bug might be related to upstream packages: Interpolation and ImageTransformation.jl

Update: downscale (e.g., Scale(0.2) ) is not affected.

[johnnychen94]

This issue might not be so easy to fix, as far as I can identify at present, the accuracy issue is due to different center generated by toaffinemap. (This might not be the root reason, but I'll just post it here as log to myself)

julia> using Augmentor: toaffinemap

julia> op = Scale(1.5);

julia> img = rand(Gray{N0f8}, 64, 64);

julia> img2 = OffsetArray(img, -1, -1);

julia> toaffinemap(op, img, (1.5, 1.5))
AffineMap([1.5 0.0; 0.0 1.5], [-16.25, -16.25])

julia> toaffinemap(op, img2, (1.5, 1.5))
AffineMap([1.5 0.0; 0.0 1.5], [-15.75, -15.75])

[Evizero]

well but numerical inaccuracy should be expected though, right? The basic idea was to always apply the affine operation around the origin, if I remember correctly.

[johnnychen94]

I guess so, but since I'm not the Augmentor expert :P, I wish I could make the test pass without relaxing them.

I'm actually thinking of a workaround to this inaccuracy, that striping the offsets, applying affine operation, and adding back offsets. But that would be tricky and there might be a performance issue as well.

[johnnychen94]

#43 does not fix this issue; it only relaxes the accuracy requirement for those identified failed test cases. Hence I'll still keep this issue open.

[johnnychen94]

Well, since general machine learning tasks don't really need such accuracy, I'll just close this issue...