ObsDimArray

[oxinabox]

One thing, I was thinking is that we could have a special alias for just specifying the observation dimentions.
Which could be useful for dispatch.

it would be:

obs_name_tuple(n) = ntuple(i->i==n ? :obs, :_)
const ObsDimArray{D, T, N} = NamedDims {obs_name_tuple(D), T, N}

const RowObsArray{T, N} = ObsDimArray{1, T, N}
const ColObsArray{T, N} = ObsDimArray{2, T, N}
const LastDimObsArray{T, N} = ObsDimArray{N, T, N}

So the following would be equivelent:

RowObsArray{Float32, 3} == NamedDimsArray{(:obs, :_, :_), Float32}

I'm not sure yet if the dispatch fully works, for mixtures of ObsDImArrays and NamedDimArrays where just one of the names is :observations, but that can potentially be changed.
(It would likely mean moving to Named being stored not as (:a,:b) type params but as Tuple{:a, :b}. Which would allow for Tuple{:obs, :<Any}. This wouild also allow for other bitvalues and types to be used as names. e,g #40 nut that is getting off-topic)

[iamed2]

Is the "observation dimension" the dimension such that a slice with all other dimensions as : and that dimension as a single value produces one observation? Sorry if I didn't phrase that well.

[oxinabox]

Correct.

[Tokazama]

This seems very similar to a lot of what https://github.com/JuliaImages/ImageAxes.jl does. Might be a helpful starting point for this.

[willtebbutt]

This would be cool to have. FWIW, I would definitely add a special case to Stheno to handle this type.

[oxinabox]

The change from the dim-names type-params being a Tuple like (:a, :b) to being a Type{Tuple} like Tuple{:a. :b}
is ugly but it facilitates dispatching on non-exact names.
Which is pretty useful.

we could also have a helper function
to construct constraitns like like
NamedTuple{<:Union{Tuple{:obs, Any}, Tuple{Any, :obs}}}
which would dispatch on having a obs dimenion in either dim1 or dim2

[nickrobinson251]

related https://github.com/invenia/ObservationDims.jl/