Improve error message for "MethodError: factornode"

[fonsp]

I created a model with a distribution that is not supported (Cauchy):

@model function lighthouse_model(observations, alpha_prior, beta_prior)
	alpha ~ alpha_prior
	beta ~ beta_prior

	for i in eachindex(observations)
		observations[i] ~ Cauchy(alpha, beta)
	end
end

The error message from infer is below, with a very long stack trace:

MethodError: no method matching factornode(::ReactiveMP.UndefinedNodeFunctionalForm, ::Type{Distributions.Cauchy}, ::Vector{Tuple{Symbol, ReactiveMP.AbstractVariable}}, ::Tuple{Vector{Int64}, Vector{Int64}})

The function `factornode` exists, but no method is defined for this combination of argument types.

Closest candidates are:
  factornode(::ReactiveMP.UndefinedNodeFunctionalForm, ::F, ::Any, ::Any) where F<:Function
   @ ReactiveMP ~/.julia/packages/ReactiveMP/1xtcb/src/nodes/predefined/delta/delta.jl:112
  factornode(::ReactiveMP.PredefinedNodeFunctionalForm, ::F, ::I, ::Any) where {F, I}
   @ ReactiveMP ~/.julia/packages/ReactiveMP/1xtcb/src/nodes/nodes.jl:190
  factornode(::F, ::I, ::Any) where {F, I}
   @ ReactiveMP ~/.julia/packages/ReactiveMP/1xtcb/src/nodes/nodes.jl:184
  ...

Improved error message

The error message can be improved! The helpful information (that I got from Dmitry) was: "The distribution Cauchy is not supported by RxInfer" and it would be nice to have this in the error message. The current error message is not helpful because I do not know why factornode was called (or what it is).

I found that Base.Experimental.register_error_hint is useful for improving MethodError messages, see eg JuliaPlots/Plots.jl#4733

Example notebook

no method factornode.html.zip

[bvdmitri]

Hey @fonsp thanks for the feedback! I made a PR ReactiveBayes/ReactiveMP.jl#431 for this