Reductions preserve the AxisArray wrapper (fixes #55)

[timholy]

Very curious result if you say @code_warntype AxisArrays.reduced_indices(axes(A), (1,)). That's why some of the @inferred are commented out.

But otherwise this seems to work well.

Oh, argh! I'm running release-0.5 and this tests on 0.5.0. This is a problem...

[timholy]

CC @tknopp.

[tknopp]

not sure if I am actually able reviewing this but my understanding is that you hook into special methods (i.e. reduced_indices) that are used in base for reductions?!

The tests look good!

[timholy]

4f1da91 may be a deal-breaker.

[timholy]

Small fix made, rebased, and inference issue filed: JuliaLang/julia#20714. I use --inline=no so frequently in debugging that I think we can't merge this PR until this issue is fixed. Since it affects julia-0.5 as well as julia-0.6, I fear it could be a long wait.

[tknopp]

@timholy: This comes up very frequently and I still do not understand what the correct workaround is. Even the type unstable form maximum(A,1) does not work anymore.

So the question is: How do I get the maximum intensity projection of a 3D AxisArray and I want the result to be still an AxisArray

[timholy]

Try this:

function fixaxes(A, axs)
    ls = map(length, indices(A))
    AxisArray(A, map((i,l)->length(i) == l ? i : i(Base.OneTo(1)), axs, ls))
end

julia> A = AxisArray(collect(reshape(1:15,3,5)), :y, :x)
2-dimensional AxisArray{Int64,2,...} with axes:
    :y, Base.OneTo(3)
    :x, Base.OneTo(5)
And data, a 3×5 Array{Int64,2}:
 1  4  7  10  13
 2  5  8  11  14
 3  6  9  12  15

julia> fixaxes(maximum(parent(A), 1), axes(A))
2-dimensional AxisArray{Int64,2,...} with axes:                                                                                                                                              
    :y, Base.OneTo(1)                                                                                                                                                                        
    :x, Base.OneTo(5)                                                                                                                                                                        
And data, a 1×5 Array{Int64,2}:                                                                                                                                                              
 3  6  9  12  15

The parent isn't strictly necessary, but if you don't want your fix to break when this gets merged you're better off including it.

[tknopp]

My issue is more fundamental

julia> typeof(B)
AxisArrays.AxisArray{Float32,4,Array{Float32,4},Tuple{AxisArrays.Axis{:x,FloatRange{Float64}},AxisArrays.Axis{:y,FloatRange{Float64}},AxisArrays.Axis{:z,FloatRange{Float64}},AxisArrays.Axis{:time,FloatRange{Float64}}}}

julia> maximum(B,1)
ERROR: BoundsError: attempt to access (:x,:y,:z,:time)
  at index [5]
 in getindex(::Tuple{Symbol,Symbol,Symbol,Symbol}, ::Int64) at ./tuple.jl:8
 in reaxis(...) at /home/knopp/.julia/v0.5/AxisArrays/src/indexing.jl:42
 in macro expansion at ./reducedim.jl:222 [inlined]
 in macro expansion at ./simdloop.jl:73 [inlined]
 in _mapreducedim!(::Base.#identity, ::Base.#scalarmax, ::Array{Float32,4}, ::AxisArrays.AxisArray{Float32,4,Array{Float32,4},Tuple{AxisArrays.Axis{:x,FloatRange{Float64}},AxisArrays.Axis{:y,FloatRange{Float64}},AxisArrays.Axis{:z,FloatRange{Float64}},AxisArrays.Axis{:time,FloatRange{Float64}}}}) at ./reducedim.jl:221
 in mapreducedim!(::Function, ::Function, ::Array{Float32,4}, ::AxisArrays.AxisArray{Float32,4,Array{Float32,4},Tuple{AxisArrays.Axis{:x,FloatRange{Float64}},AxisArrays.Axis{:y,FloatRange{Float64}},AxisArrays.Axis{:z,FloatRange{Float64}},AxisArrays.Axis{:time,FloatRange{Float64}}}}) at ./reducedim.jl:237
 in mapreducedim(::Function, ::Function, ::AxisArrays.AxisArray{Float32,4,Array{Float32,4},Tuple{AxisArrays.Axis{:x,FloatRange{Float64}},AxisArrays.Axis{:y,FloatRange{Float64}},AxisArrays.Axis{:z,FloatRange{Float64}},AxisArrays.Axis{:time,FloatRange{Float64}}}}, ::Int64) at ./reducedim.jl:268
 in maximum(::AxisArrays.AxisArray{Float32,4,Array{Float32,4},Tuple{AxisArrays.Axis{:x,FloatRange{Float64}},AxisArrays.Axis{:y,FloatRange{Float64}},AxisArrays.Axis{:z,FloatRange{Float64}},AxisArrays.Axis{:time,FloatRange{Float64}}}}, ::Int64) at ./reducedim.jl:320

[timholy]

If you do that to parent(B) you won't have that problem; my suggested approach should work around that dandily.

That said, I didn't realize that was the problem. Maybe the same as #62? I'd forgotten about that one.

[tknopp]

Sorry for mixing things. As you said its two independent issues. But from a higher level its simply a regression on the previous Images.jl workflow. The parent thing does not work. But I open a dedicated issue with testcase.

[tkelman]

If this is something that differs between 0.5.0 and 0.5.1, is it worth testing 0.5.0 as a separate matrix entry on travis, given not everyone has upgraded right away to 0.5.1?

[timholy]

I didn't realize you do could that. See https://travis-ci.org/JuliaArrays/AxisArrays.jl/builds/215065371. I think I worked around the problem in a later commit, so I edited the comment above.

(Given that it's passing on 0.5.0, I don't think it's worth adding to the test matrix, but others might feel differently.)