Interpolate

[DanDeepPhase]

Any interest in supporting interpolation along the axes?

I was able to hack it in my code with something like:

using Interpolations, DimensionalData
import Interpolations.interpolate

interpolate(da::DimArray) = interpolate(Array.(dims(da)), da, Gridded(Linear())) 

xs = ys = 1:10
zs = [rand() for x in xs, y in ys]
M = DimArray(zs, (Y(ys),X(xs)))

itp = interpolate(M)
itp[3.3, 5.1]
[itp[y,x] for y in 3.1:0.35:7.2, x in 2:4] 

I saw you use something in rasters for re-gridding, which i think this approach is best suited for. some tweaks would be needed to get reverse ordered vectors working.

There may be a more intrusive / better way to do this by extending or creating a variant of At()

[rafaqz]

Rasters users gdalwarp because it handles projections and has a lot of options. But native interpolation is better eventually.

Unfortunately Interpolations.jl is a very heavy dependency, and this package is a dep of other packages so we have to keep things lean.

DimensionalDataInterpolations.jl could be useful.

[rafaqz]

Revisiting this: with Julia 1.9 and this PR JuliaLang/julia#47695 we can add a weak dependency on Interpolations.jl.

If you want to make a PR that extends Intpolations.jl methods to work on AbstractDimArray in a weak dependency in this package, that would be a very useful addition.

Edit: also note that with Regular sampling you do not need to use Gridded interpolation - but you do if any LookupArray for the axes has Irregular sampling. We can really customize the interpolation method to match the traits of the LookupArray.

[kapple19]

I'm happy to look into implementing DataInterpolationsND.jl as an extension.

[rafaqz]

Yeah, that would be great honestly.

The tricky part will be the interface, like how do we let users specify interpolation dimension types without having to define custom syntax for it all.

[kapple19]

I might approach this some small PRs at a time, it's a bigger task than I anticipated, and I want more features the more I think about it 😅

Like, DimStack and DimTree support!

But will get the simple things through first, like dispatching correctly to DataInterpolations.jl (1D) or DataInterpolationsND.jl.

Is it acceptable to parent the array and dims in the call to the interpolation constructors? Or is there a better way?

Also last I checked, the interpolator calles accept vector inputs for vector outputs. We want that too?

[rafaqz]

Is it acceptable to parent the array and dims

Not sure what you mean exactly but for most base external methods we normally do something with dims that sets up the args or kwargs, then pass parent(da) to dispatch to the main methods, as the main argument.

Probably we want everything that can work in the package to work here too. But I'll have to see a PR to really know.

As for small PRs, they will all have to offer some useful working functionality to be worth merging, but yes otherwise that's fine.

[rafaqz]

Oh its also ok to document this as new and breakable without a breaking bump (like DimTree is) so theres less pressure to get it right straight away, but people get to use it at least and start giving feedback

[kapple19]

DimensionalDataInterpolationsExt.jl 😂