Pullback in constructor rrule should unwrap a NamedDimsArray gradient back to an Array

[mzgubic]

From @mcabbott: #178 (comment)

I am too slow but shouldn't this remove names from the array it gets? On the assumption that similar or broadcasting in the computation of a gradient for the array with names may, probably should, produce another array with names?

julia> nda = NamedDimsArray(rand(Int8,3,4), (:x, :y));

julia> gradient(x -> sum(x), nda)  # does not know about names
(Fill(1, 3, 4),)

julia> gradient(x -> sum(x; dims=1)[1], nda)  # involves broadcasting, so might? 
ERROR: DimensionMismatch("destination axes (Base.OneTo(3), Base.OneTo(4)) are not compatible with source axes (Base.OneTo(1), Base.OneTo(4))")

julia> gradient(x -> sum(x; dims=1)[1], nda.data)
(Int8[1 0 0 0; 1 0 0 0; 1 0 0 0],)

[mcabbott]

The error in my example seems to have gone away. But better examples would be:

julia> using NamedDims, Zygote

julia> nda = NamedDimsArray(rand(Int8,3,4), (:x, :y));

julia> gradient(x -> sum(NamedDimsArray(x, (:x, :y)); dims=1)[1], rand(3,4))[1]
3×4 NamedDimsArray{(:x, :y), Float64, 2, Matrix{Float64}}:
 1.0  0.0  0.0  0.0
 1.0  0.0  0.0  0.0
 1.0  0.0  0.0  0.0

julia> gradient(x -> sum(NamedDimsArray(x, (:x, :y)) * nda'), rand(3,4))[1]
3×4 NamedDimsArray{(:_, :y), Float64, 2, Matrix{Float64}}:
 94.0  121.0  -116.0  -165.0
 94.0  121.0  -116.0  -165.0
 94.0  121.0  -116.0  -165.0

[oxinabox]

Tangent should have same names as primal, no?

[mcabbott]

Yes, ideally. Many rules won't do that, but ProjectTo could enforce it.

[oxinabox]

I do not understand the title of this issue.