Indexing by name when multiple dimensions have same name

[nickrobinson251]

We currently consider this undefined behaviour -- and i wonder if that's an unfortunate middle ground.

In practice (using v0.2) we have consistent and reasonable behaviour: when there are duplciate names. indexing by name indexes both dimenstions

nda = NamedDimsArray{(:a, :b, :b)}(rand(10, 20, 30));
nda[a=2, b=3] == nda[2, 3, 3]
getindex(nda, b=1) == nda[:, 1, 1]

I wonder if we should commit to supporting this (and document it)?

Alternatively, we could go the other direction - perhaps we think this could become a source of subtle bugs for users - and error upon "ambiguous" indexing, i.e. with naems(nda) == (:a, :b, :b), should nda[a=2, b=3] throw an informative error?

[nickrobinson251]

The more I think about this, the more I like it, and think we should document it. Although I'm still not sure I would want to actually use this behaviour for anything serious... it's a bit magic.

[mcabbott]

Having the same name twice seems fairly natural for d = Diagonal(NamedDimsArray(rand(3), :i)), for which d[i=1] is currently an error, but could be a number.

The other place this seems likely to occur is an outer product, like so:

julia> m = NamedDimsArray(rand(2,4), (:i, :k));

julia> p = m' * m
4×4 NamedDimsArray{(:k, :k),Float64,2,Array{Float64,2}}:

julia> p[k=3]
0.22535458930936356

Indexing that down the diagonal seems wrong to me.

One idea might be to forbid the creation of such objects, and insist that * produce NamedDimsArray{(:k, :k′), here. Would this cause other problems?

(Presumably creation by hand like NamedDimsArray(ones(2,2), (:k, :k)) should then be an error.)

[nickrobinson251]

Having the same name twice seems fairly natural for d = Diagonal(NamedDimsArray(rand(3), :i)), for which d[i=1] is currently an error, but could be a number.

this is an error becuase you cannot index a Diagonal by name, only a NamedDimsArray (which you know #65, but in case other folks don't)

One idea might be to forbid the creation of such objects, and insist that * produce NamedDimsArray{(:k, :k′), here. Would this cause other problems?

It is definitely an aim of this package to support duplicate names. If that's not written down publicly yet, then it should be (either in the readme or when we add docs #30) :) e.g. we want behaviour like

julia> using Statistics, NamedDims

julia> nda = NamedDimsArray(rand(4, 2), (:x, :y));

julia> Σ = cov(nda)
2×2 NamedDimsArray{(:y, :y),Float64,2,Array{Float64,2}}:
  0.0258206   -0.00923985
 -0.00923985   0.0569803

The only question is whether we should officially support Σ[y=1] == Σ[1, 1]

[mcabbott]

It is definitely an aim of this package to support duplicate names.

OK. Then I would vote that p[k=3] should be an error, it seems too surprising to return a number when indexing a full matrix that accidentally has duplicate names. Although I see that Σ[y=1] makes some sense.

(Perhaps things like NamedDims{...Diagonal} should evade the error, by not really being two dimensional.)