Can't @gg user-defined functions

[cscherrer]

Hi @thautwarm, I had mentioned this issue is the Soss issues, but I think I wasn't very specific about it.

Everything works great for models like this:

julia> m1 = @model begin
           x ~ Normal()
       end;

julia> rand(m1())
(x = -0.11329202445030989,)

Calling rand in this case passes the following to @gg:

julia> sourceRand()(m1)
quote
    x = rand(Normal())
    (x = x,)
end

I think it will be reasonably common for users to want to define their own distributions, and I'd hate for this always to require a PR to Distributions or Soss. So maybe they have a simple case like this:

julia> struct N end

julia> Distributions.rand(::N) = rand(Normal())

julia> rand(N())
0.02887276952144644

So far so good. But when they try to use it in a model, ...

julia> m2 = @model begin
           x ~ N()
       end;

julia> rand(m2())
ERROR: UndefVarError: N not defined
Stacktrace:
 [1] getproperty at ./Base.jl:13 [inlined]
 [2] macro expansion at /home/chad/.julia/packages/GeneralizedGenerated/x3uMp/src/closure_conv.jl:155 [inlined]
 [3] _rand(::Model{NamedTuple{(),T} where T<:Tuple,begin
    x ~ N()
end}, ::NamedTuple{(),Tuple{}}) at /home/chad/.julia/packages/GeneralizedGenerated/x3uMp/src/closure_conv.jl:155
 [4] rand(::Soss.JointDistribution{NamedTuple{(),Tuple{}},NamedTuple{(),T} where T<:Tuple,begin
    x ~ N()
end}) at /home/chad/git/jl/Soss/src/rand.jl:8
 [5] top-level scope at REPL[68]:1

I think this is because the source code is like this:

julia> sourceRand()(m2)
quote
    x = rand(N())
    (x = x,)
end

The generated function docs say users should not attempt...

Calling any function that is defined after the body of the generated function. This condition is relaxed for incrementally-loaded precompiled modules to allow calling any function in the module.

So my guess is this might be the root of the problem.

Some possibilities I can imagine (which might not all help or be feasible):

1. Require everything user-defined to be passed as an argument. This currently works, but is kind of unwieldy.
2. Allow $ interpolation. I think this could work, but it does leave the Soss-provided distributions feeling kind of magic, so anything a user builds is second-class.
3. I had thought there could be a trick with the evaluation order, like defining N before using Soss. This doesn't help, but could there be some variant of this that will work?
4. Ideally (though I have no idea if it's possible or how much work would be involved) it would be pretty great for @gg (or some variant of it) to allow anything in scope at call-time to be referenced.

Any ideas?

[thautwarm]

Hi Chad, the problem has been detected.

The use of @gg in Soss is in Soss module, while the definition of N is in the Main module, thus
the reference of N failed in your case.

[thautwarm]

I'll make a PR to refine the use of @gg in Soss. In fact, we should store the current module when invoking @model.

[thautwarm]

It seems to be really difficult now to support the expected feature without using $.

Once the generalized functions get defined, its global scope is decided as well, thus, according to Soss/rand.jl, the module to look up global variables is specified to be Soss.

However, the best way IMHO to achieve the expected behaviour is,

• Soss.jl side: adding an extra field to Soss.Model to store the model's definition module.
• GG.jl side: introducing a new @gg macro to support passing in the module to look up global variables: @gg @__MODULE__ f(x) = N(x) # now N is from @__MODULE__

[cscherrer]

Fixed in #28