For(...) do i ... broken using Soss with or later about soss.jl

[cscherrer]

Here's a MWE, removed from Soss:

julia> using Soss, MeasureTheory, GeneralizedGenerated

julia> g = GG.mk_function(:(x -> Normal(x,1)))
function = (x;) -> begin
    begin
        (GeneralizedGenerated).Normal(x, 1)
    end
end

julia> For(g, 3)
ERROR: MethodError: no method matching productmeasure(::ggfunc-function, ::UnitRange{Int64})
Closest candidates are:
  productmeasure(::Any, ::Any, ::Any) at /home/chad/git/MeasureBase/src/combinators/smart-constructors.jl:69
  productmeasure(::Kernel, ::Any) at /home/chad/git/MeasureBase/src/combinators/smart-constructors.jl:75
  productmeasure(::Returns, ::Any, ::Any) at /home/chad/git/MeasureBase/src/combinators/smart-constructors.jl:82
  ...
Stacktrace:
 [1] For(f::ggfunc-function, n::Int64)
   @ MeasureBase ~/git/MeasureBase/src/combinators/for.jl:82
 [2] top-level scope
   @ REPL[53]:1

mk_function calls GeneralizedGenerated.closure_conv, which produces

julia> typeof(g)
ggfunc-function

julia> typeof(g) |> supertype
Any

So the issue is that MeasureBase.productMeasure doesn't have a generic method, but wants a <:Function, and a ggfun-function doesn't meet this criteria.

The simplest fix for this is to add a method

productmeasure(f, pars) = productmeasure(f, identity, pars)

For example, this works:

julia> MeasureBase.productmeasure(f, pars) = MeasureBase.productmeasure(f, identity, pars)

julia> m = @model begin 
           x ~ For(3) do j Normal(j,1) end
       end;

julia> rand(m())
(x = [1.9296440232404293, 3.1574221730525336, 4.1352330633979255],)

But it turns out this just kicks the can down the road:

julia> logpdf(m(), rand(m()))
ERROR: MethodError: no method matching basekernel(::ggfunc-function)
Closest candidates are:
  basekernel(::Kernel) at /home/chad/git/MeasureBase/src/kernel.jl:73
  basekernel(::Returns) at /home/chad/git/MeasureBase/src/kernel.jl:74
  basekernel(::Function) at /home/chad/git/MeasureBase/src/kernel.jl:71

To be continued...

[Tuebel]

Thanks for the quick fix!

[cscherrer]

Thanks for letting me know about this! This is a high priority to fix, obviously. I think (hope?) it should be a quick fix, then we should add tests to be sure to catch this going forward. I'll spend some time on it today.

[cscherrer]

Hi @Tuebel , thanks again for this issue. There's a new release with tests that include

    @testset "https://github.com/cscherrer/Soss.jl/issues/305" begin
        m = @model begin 
            x ~ For(3) do j Normal(μ=j) end
        end;
        
        @test logpdf(m(), rand(m())) isa Float64
    end