Use of arbitrary distributions from Distributions.jl

[bartvanerp]

As discussed, it should be possible for users to use arbitrary distributions from Distributions.jl, without them needing to be defined in RxInfer.jl. Here just a reminder: As shown, this will require some lines of code, once the new GraphPPL version gets integrated in RxInfer. Would be great if the interface names get automatically adopted from the fieldnames in Distributions.jl.

[wouterwln]

Note to self:

GraphPPL.NodeBehaviour(::Type{<:Distributions.Distribution}) = GraphPPL.Stochastic()
GraphPPL.rhs_to_named_tuple(::GraphPPL.Atomic, t::Type{<:Distributions.Distribution}, interface_values) = NamedTuple{fieldnames(t)}(interface_values)
GraphPPL.interfaces(t::Type{<:Distributions.Distribution}, ::Val{3}) = (:out, fieldnames(t)...)

These lines implement whatever is necessary in order to create arbitrary nodes from Distributions.jl objects. Preferably we put this in RxInfer.jl since I don't want to take a dependency on Distributions.jl in GraphPPL.jl. This also gives some weird error if yo do Normal(0, 1, 1) or any other illegal stuff which we should handle appropriately but at least it provides a MWE in order to make arbitrary distributions possible with the desired interface names.

[bvdmitri]

We definitely put this either in RxInfer or in an extension (if Distributions is loaded by a user, then we could define extra functions, this is new in 1.9)

[wouterwln]

I merged a PR that should implement the extension. @bvdmitri are you aware of any way in which I can test this or should we leave it be?

[wouterwln]

There are slight issues with this, as not all distributions from Distributions.jl have their interfaces as their struct fields (for example, a Dirichlet distribution has the following interfaces:

julia> GraphPPL.interfaces(Dirichlet, 1)
(:out, :alpha, :alpha0, :lmnB)

We can implement these "special" cases in the extension as well, but that kind of defies the point of this extension. Any suggestions?

[bvdmitri]

@wouterwln I think you can load Distributions in the tests and try to create an arbitrary graph with some arbitrary distribution and test that interfaces match your expectations.

With the second issue about differences in the struct fields and desired interfaces it will happen more often I believe. You should look into the params method, but I'm not sure how to get actual names. That might require some manual work indeed.

[wouterwln]

This is as fixed as it's going to be. I'm going to close this to clean up the overview of open issues