Slicing with colon produces Vector instead of ComponentVector

[baggepinnen]

I'm trying to get ComponentVector to work with Optim.ParticleSwarm and am facing the following issue: when slicing into a matrix where there are components along the first axis, the result is a regular vector

julia> A
4×10 ComponentMatrix{Float64} with axes Axis(a = 1:2, b = 3:4) × FlatAxis()
 0.0           0.0           0.0           …  0.0           0.0           0.0
 0.0           0.0           0.0              0.0           0.0           0.0
 6.92009e-310  6.92009e-310  6.92009e-310     6.92009e-310  6.92009e-310  6.92017e-310
 6.92009e-310  6.92009e-310  6.92009e-310     6.92009e-310  1.0e-323      6.92017e-310

julia> A[:,1]
4-element Vector{Float64}:
 0.0
 0.0
 6.92009e-310
 6.92009e-310

In this case, I believe it should be possible to return a ComponentVector even if I index the component dimension with :?

[jonniedie]

Yep, it should do that. This is a bug that must have been introduced when I recently changed the flat (non-symbolic) indexing behavior.

Although this highlights an issue that I’ve had since the beginning of ComponentArrays: Should non-symbolic indexing produce a view or a new array? Since named views are kinda the whole purpose of ComponentArrays, I always had it set up so all indexing creates a view. But this led to some bugs when interacting with other packages because the general expectation is that indexing an array in the “normal” way should always produce a new array. So I recently changed it so that indexing would only create views when symbolic indices are used. But this puts A[:,1] in a weird spot: It feels to me like ComponentArray -> ComponentArray operations like this should always use views, but that’s wrong for the above reasons and would continue to introduce bugs. So when I push this fix, it will be such that @view A[:,1] is needed if you want the result to be a view.

[baggepinnen]

I was expecting A[:,1] to return a copy, but with the type being a ComponentVector, so requiring @view A[:,1] to obtain a view feels in line with my expectations.

[baggepinnen]

Thanks for the fix! The PR to Optim is here
JuliaNLSolvers/Optim.jl#883
I provided a naive implementation of

Optim.similar_axis(x::ComponentArray, n) = x .* zeros(length(x), n)

which gives me

ComponentMatrix{Float64} with axes Axis(a = 1:2, b = 3:4) × FlatAxis()

but I'm sure there's a better way?