Very unintuitive error message when the syntax is invalid

[ForceBru]

Code:

begin
	import Pkg
	Pkg.activate(temp=true)
	Pkg.add(name="Turing", version="0.16.0", io=devnull)
	Pkg.add(name="Distributions", io=devnull)
	Pkg.status()
end


using Turing, Distributions

@model function(data::AbstractVector)	
	p ~ Dirichlet(2, 1.0)
	μ1 ~ Normal(-4e-4, 2.1e-4)
	μ2 ~ Normal( 4e-4, 2.1e-4)
	
	Σ ~ filldist(Uniform(0, 2.1e-4), 2)
	
	data ~ filldist(
		UnivariateGMM([μ1, μ2], Σ, p),
		length(data)
	)
end

Error:

~/test $ julia turing_bug2.jl
  Activating new environment at `/var/folders/ys/3h0gnqns4b98zb66_vl_m35m0000gn/T/jl_nVue8d/Project.toml`
      Status `/private/var/folders/ys/3h0gnqns4b98zb66_vl_m35m0000gn/T/jl_nVue8d/Project.toml`
  [31c24e10] Distributions v0.25.2
  [fce5fe82] Turing v0.16.0
ERROR: LoadError: LoadError: KeyError: key :name not found
Stacktrace:
 [1] getindex(h::Dict{Symbol, Any}, key::Symbol)
   @ Base ./dict.jl:482
 [2] build_output(modelinfo::Dict{Symbol, Any}, linenumbernode::LineNumberNode)
   @ DynamicPPL ~/.julia/packages/DynamicPPL/emAbx/src/compiler.jl:385
 [3] model(mod::Module, linenumbernode::LineNumberNode, expr::Expr, warn::Bool)
   @ DynamicPPL ~/.julia/packages/DynamicPPL/emAbx/src/compiler.jl:91
 [4] var"@model"(__source__::LineNumberNode, __module__::Module, expr::Any, warn::Any) (repeats 2 times)
   @ DynamicPPL ~/.julia/packages/DynamicPPL/emAbx/src/compiler.jl:82
in expression starting at /Users/forcebru/test/turing_bug2.jl:12
in expression starting at /Users/forcebru/test/turing_bug2.jl:12
~/test [1] $

---

I literally spent 15 minutes trying to understand what was wrong. Turns out, I was using too much R lately, and @model function(data::AbstractVector) should've been @model function model_name(data::AbstractVector).

However:

1. The error message KeyError: key :name not found looks like an internal error that doesn't tell me what the problem is.
2. function(arg) arg + 1 end is actually valid Julia syntax, so one can write my_func = function(arg) arg + 1 end, much like in R. I guess Turing's models aren't really functions (even though they look exactly like functions), so it's probably fine that some syntax isn't supported, but the error message could use some improvement, IMO

[devmotion]

Oh yeah, this error message is quite unintuitive. Basically, @model can only work with functions that have a name and function(arg) ... only creates an anonymous function (similar to x -> x + 1). I am not even sure if MacroTools.splitdef (https://fluxml.ai/MacroTools.jl/stable/utilities/#Function-definitions-1) can deal with these function definitions at all. A simple fix would be to just check if the function has a name by explicitly checking if the dictionary returned by MacroTools.splitdef returns a :name.

Regarding 2., actually the main point of the @model macro is that it works with regular Julia functions and modifies the user-provided function only very slightly by adding additional arguments for passing variables and changing the behaviour of executing it and replacing the tilde expressions. The only point is that the function requires a name 😄 There are some more explanations here: https://turing.ml/dev/docs/for-developers/compiler