Splitting multidimensional array with getindex()

[takagiy]

map or broadcasting syntax takes multidimensional arrays as just 1 dimensional sequences. mapslices always tries to return an array of the same number of dimensions as the taken array, this sometimes causes DimensionMismatch or make a virtual 1 dimensional array extending along a strange axis. There is no specific way to map a function taking an array on a multidimensional array.

a = zeros(2, 2, 5)

"some processing that takes one 2d array and transform it to a 3d array"
someprocess(a::AbstractArray{<:Any, 2}) = rand(3, 3, 3)

map(someprocess, a)
#ERROR: MethodError: no method matching someprocess(::Float64)
mapslices(someprocess, a, [1, 2])
#ERROR: DimensionMismatch("tried to assign 3 elements to 2 destinations")

Another range expression that acts like a dial knob(hereinafter, i refer to it as --) may manage it.(Maybe the new subtype of AbstractArray that represents the split multidimensional array will be needed for the return value type of getindex)

a[:, :, --]
# the return value behaves like [a[:, :, 1], a[:, :, 2], a[:, :, 3], a[:, :, 4], a[:, :, 5]]. (or simply is it)
a[:, :, 2--3]
# the return value behaves like [a[:, :, 2], a[:, :, 3]]. (or simply is it)

map(someprocess, a[:, :, --]) # and
someprocess.(a[:, :, 2--3]) # will be available

[mbauman]

Hi @takagiy — welcome! You may be interested in the JuliennedArrays package. It takes an array and returns an iterable that yields slices of it. Instead of --, it uses * to specify the "dial knob."

using JuliennedArrays
map(someprocess, julienne(Views, a, (:, :, *)))
someprocess.(julienne(Views, a, (:, :, *)))

I'm going to close this in favor of some of the existing discussions like #3893 and #23645 to make sure our thoughts on the subject don't get too scattered.