Issue: Misleading "RuleMethodError" exception leading to misleading "inference_postprocess" error (read below)

[ofSingularMind]

I developed a custom node, with custom update rules and custom meta function. I then duplicated this node with a small adjustment, but forgot to update my @meta function meta_fn to provide a meta function for this duplicate-altered node. The below describes the misleading debug signals I received while tracing and fixing this problem.

Because my meta-function was wrong, connected variables were not getting updated within RxInfer.batch_inference() function calls (presumably because the update rules require a meta function that was unavailable).

Within the

for (key, value) in fdata
    update!(cacheddatavars[key], get_data(value))
end

function calls inside RxInfer.batch_inference(), and utilizing VS Code debugger set to break on All Exceptions, I would receive notice of an exception, a RuleMethodError, at the following code in ReactiveMP.message.jl.

ruleoutput = rule(ruleargs...)
# if `@rule` is not defined, the default behaviour is to return 
# the `RuleMethodError` object
if ruleoutput isa RuleMethodError
    !isnothing(mapping.rulefallback) ? mapping.rulefallback(ruleargs...) : throw(ruleoutput)
else
    ruleoutput
end

The attached image shows the exception along with the arguments passed for an appropriate update rule. We can notice two things. First, the suggested new rule that might resolve the issue is already exactly included in the list of available rules. Second, when we look at the ruleargs, we see the passed variable exactly match these type specifications. So, this was very confusing. But what the exception does not list is the missing meta function. The exception text is included in exception.txt.

After this exception is softly thrown, it would be caught within the RxInfer.batch_inference() try ... catch routine to fill out a variable potential_error below.

try 
  ...
catch error
  potential_error = inference_process_error(error, !catch_exception)
end

So, this potential_error variable would be filled and then code execution would continue a few lines until the following where posterior inference values are to be post-processed. Because the variables were not being updated from inference, the inference_postprocess call would fail because it cannot process nothing values.

  posterior_values = Dict(variable => inference_postprocess(postprocess, getvalues(actor)) for (variable, actor) in pairs(actors_rv))

This is where RxInfer would quit, throwing the error shown in error.txt, summarized as

MethodError: no method matching inference_postprocess(::DefaultPostprocess, ::Nothing)
The function `inference_postprocess` exists, but no method is defined for this combination of argument types.

So, the final fix was to change my meta function from:

@meta function meta_gp_regression(Xu, Ψ0, Ψ1_trans, Ψ2, KuuL, Uv)
    UniSGP_AMF() -> UniSGP_AMFMeta(nothing, mean_fn, Xu, Ψ0, Ψ1_trans, Ψ2, KuuL, kernel, Uv, 0, 0)
end

to

@meta function meta_gp_regression(Xu, Ψ0, Ψ1_trans, Ψ2, KuuL, Uv)
    UniSGP_AMF() -> UniSGP_AMFMeta(nothing, mean_fn, Xu, Ψ0, Ψ1_trans, Ψ2, KuuL, kernel, Uv, 0, 0)
    UniSGP_AMF_Grad() -> UniSGP_AMFMeta(nothing, mean_fn, Xu, Ψ0, Ψ1_trans, Ψ2, KuuL, kernel, Uv, 0, 0)
end

where UniSGP_AMF_Grad is the duplicate-altered function, and this presumably allowed the variables to now find an appropriate update rule to use and all was fixed :)

I guess I just don't feel this was the error I wanted to receive for this issue with my code. Maybe there's a better option.

Thanks! Alex

[wouterwln]

@ofSingularMind so if I understand correctly, everything would've been good if the RuleMethodError would have the "expected" meta specification in the error message? By the way, +1 on this issue, I have been caught off guard by this quite often. If this is indeed what you mean, let's fix it

[ofSingularMind]

Not quite. When I run infer(), the error I get is the no method matching inference_postprocess(...) one, not the RuleMethodError one. So, I presume I should get the RuleMethodError one, and it should be referencing the meta as part of the required rule specifications. Then, if we look back at the picture, we would see ruleargs[8] (the meta spec) is equal to nothing, for which there is no rule :). Thanks :)

[ofSingularMind]

Actually, now that I think about it. The suggestion in the exception is exactly right. It is asking that I define that same rule that I already have, except without a meta specification, which then my ruleargs satisfy. So.. I'm not sure the solution.

[ofSingularMind]

The last comment I have then is that none of those "already existing rules" actually exist. Not if they are implying "no meta specification."

[bvdmitri]

oops ReactiveBayes/ReactiveMP.jl#468