Type stability regression for mean(::Array{RGB{N0f8}}) from 0.6 to 0.7

[Evizero]

I don't know where the isssue was introduced exactly (or where this issue should be opened), but it seems to occur in Base._mapreduce(identity, Base.add_sum, ...) somewhere.

0.6

julia> using Colors, ColorVectorSpace, FixedPointNumbers

julia> A = rand(RGB{N0f8}, 2, 2)
2×2 Array{ColorTypes.RGB{FixedPointNumbers.Normed{UInt8,8}},2}:
 RGB{N0f8}(0.71,0.686,0.294)   RGB{N0f8}(0.871,0.322,0.667)
 RGB{N0f8}(0.369,0.055,0.478)  RGB{N0f8}(0.016,0.035,0.533)

julia> Test.@inferred mean(A)
RGB{Float64}(0.4911764705882353,0.27450980392156865,0.4931372549019607)

0.7

julia> using Colors, ColorVectorSpace, FixedPointNumbers, Statistics, Test

julia> A = rand(RGB{N0f8}, 2, 2)
2×2 Array{RGB{N0f8},2} with eltype RGB{Normed{UInt8,8}}:
 RGB{N0f8}(0.188,0.757,0.051)  RGB{N0f8}(0.922,0.871,0.012)
 RGB{N0f8}(0.62,0.184,0.69)    RGB{N0f8}(0.863,0.498,0.675)

julia> Test.@inferred mean(A)
ERROR: return type RGB{Float64} does not match inferred return type Union{RGB{Float32}, RGB{Float64}}
Stacktrace:
 [1] error(::String) at ./error.jl:33
 [2] top-level scope at none:0

[kimikage]

Let us sort out the issue.

First of all, Colors.jl does not provide arithmetic operation methods for colors except XYZ and LMS. (cf. Support arithmetic only for linear color spaces)
So, it seems to me that the inferred return type Union{RGB{Float32}, RGB{Float64}} comes from the codes of ColorVectorSpace.jl.

Second, what is the expected return type of mean(::Array{RGB{N0f8}})?
If you mind the type stability because of the processing speed, I wonder whether RGB{Float64} is always best. Moreover, to stop using mean may be a possible measure.
If the problem is not the speed, why not explicitly convert the result into preferred type? For example,

julia> Test.@inferred convert(eltype(A), mean(A))

By the way, the return type of Colors.weighted_color_mean() is the same as the inputs, i.e. RGB{N0f8} in this case, although it may cause a precision issue.

Well, if it is a pure backward compatibility issue, it's time to give up.
Julia v0.7 (v1.0) includes many breaking changes, you know.

[timholy]

@Evizero, sorry that I failed to notice this in the crush of updating packages to Julia 1. Fixed over at JuliaGraphics/ColorVectorSpace.jl#103