Migrate @relation to use the ADT approach

[jpfairbanks]

Currently this is partially implemented in SyntacticModels. We should backport this to Catlab.

[lukem12345]

@jpfairbanks Would the closing PR take care of the use case of programmatic construction?

There is code in DiagrammaticEquations.jl that has to do metaprogramming to use the constructor. But it would be great if equivalent functionality was offered upstream.

[jpfairbanks]

The code you linked could just use the programmatic interface for the ACSet as it already is.

I haven't test this, but it should be pretty close.

using Catlab
using Catlab.Programs
using Catlab.RelationalPrograms
using Catlab.WiringDiagrams.Undirected

function construct_relation_diagram(boxes::Vector{Symbol}, junctions::Vector{Vector{Symbol}})
  d = UntyledRelationDiagram{Symbol, Symbol}()
  juncs = unique(flatten(junctions))
  map(juncs) do j
    add_junctions!(d, variable=j)
  end
  map(boxes, junctions) do b, js
    # Expr(:call, b, j...)
    box = add_box!(d, b, js...)
    ports = incident(d, box, :port)
    map(ports, js) do p, j
      set_junction!(d, p, j)
    end
  end
  return d
  # quote @relation () begin $(tables...) end end |> eval
end

[jpfairbanks]

Ok, I also looked at what I had in SyntacticModels, and it is a little more complicated to get the outer ports and variable types right.

https://github.com/AlgebraicJulia/SyntacticModels.jl/blob/main/src/uwd.jl#L158

Yes, once this refactor is implemented you should have an easier time writing constructors like this that don't use macros because you can build a UWDExpr programmatically, rather than constructing the julia expression.

[lukem12345]

Yes as far as I can remember, there is another potential footgun due to the fact that the mapping from global ports to local ports depends on the order of rows in the Ports table. This came up in the context of implementing oapply for Decapodes (which ultimately avoided the issue due to not needing to make use of outer ports.) I think that code that needs that info is essentially relying on an “implementation detail” kind of behavior, though the lines get blurry.

[jpfairbanks]

@samuelcohen1 Evan says that there are good examples of this in https://github.com/AlgebraicJulia/DataMigrations.jl/blob/main/src/DiagrammaticPrograms.jl