Different data structures for load in mac-os and linux

[kirnap]

When I load an image from the same source I got different data types in mac and linux, Namely,
Mac version:

julia> arr = load(img);
julia> typeof(arr)
Array{ColorTypes.RGB4{FixedPointNumbers.Normed{UInt8,8}},2}

Linux version:

julia> arr = load(img);
julia> typeof(arr)
Array{ColorTypes.RGB{FixedPointNumbers.Normed{UInt8,8}},2}

That difference creates a problem in permutedims function. Here is an example:

julia> typeof(c1)
ImageCore.ChannelView{FixedPointNumbers.Normed{UInt8,8},3,Array{ColorTypes.RGB4{FixedPointNumbers.Normed{UInt8,8}},2}}
julia> permutedims(c1, (3,2,1))
ERROR: DimensionMismatch("new array has 224 color channels, must have 3")
Stacktrace:
 [1] check_ncolorchan(::Array{ColorTypes.RGB4{FixedPointNumbers.Normed{UInt8,8}},2}, ::Tuple{Int64,Int64,Int64}) at /Users/omer/.julia/v0.6/ImageCore/src/colorchannels.jl:356
 [2] similar(::ImageCore.ChannelView{FixedPointNumbers.Normed{UInt8,8},3,Array{ColorTypes.RGB4{FixedPointNumbers.Normed{UInt8,8}},2}}, ::Type{FixedPointNumbers.Normed{UInt8,8}}, ::Tuple{Int64,Int64,Int64}) at /Users/omer/.julia/v0.6/ImageCore/src/colorchannels.jl:117
 [3] permutedims(::ImageCore.ChannelView{FixedPointNumbers.Normed{UInt8,8},3,Array{ColorTypes.RGB4{FixedPointNumbers.Normed{UInt8,8}},2}}, ::Tuple{Int64,Int64,Int64}) at ./permuteddimsarray.jl:116

here is my ad-hoc solution...

macfix = convert(Array{FixedPointNumbers.Normed{UInt8,8},3},  c1);
julia> permutedims(macfix, (3,2,1)); # works perfectly fine

[timholy]

The different structures are a consequence of what the underlying C libraries return.

The issue about permutedims is definitely "our" problem, and it's a really awful bug. Same as JuliaImages/ImageCore.jl#48. I have a rewrite for 0.7 here that should fix this, and if you're brave you could test it.

Fixing this on 0.6 is unfortunately a breaking change, and consequently I'm inclined to leave it like it is and do all my breaking with the upgrade to julia 0.7.

[rsrock]

Ugh, RGB4 vs RGB again. Will #52 affect QuartzImageIO when it lands? I should make note of that.

[timholy]

I posted a much longer explanation of what's happening in JuliaImages/ImageCore.jl#52. The bottom line is that if you were formerly using reinterpret, you're likely to have to change it to reinterpretc. If you were using colorview, then I'd guess there will be no code changes required.

If your tests explicitly tested the type of object returned from load operations, you'll likely have to change those. That was actually the most annoying part of JuliaImages/ImageCore.jl#52, I had written a lot of such tests that I later came to regret. I switched most of them to just checking the eltype and dimensions.

[johnnychen94]

I haven't observed any similar issues since Julia v1.0, so I guess this issue is already fixed and thus become outdated?