Initialization macro lacks features compared to model macro

[bartvanerp]

There are differences between the @initialization macro and the @model macro, which might be confusing for users. An example:

@model function model() 
    x ~ NormalMeanPrecision(0.0, 1.0) # allowed
    x ~ Normal(mean = 0.0, precision = 1.0) # allowed
end

@initialization begin
    q(x) = Normal(mean = 0.0, precision = 1.0) # not allowed
    q(x) = NormalMeanPrecision(0.0, 1.0) # allowed
end

Would be nice if we can streamline this.

[bvdmitri]

The issue here is that the code that transforms Normal(mean = 0.0, precision = 1.0) to NormalMeanPrecision(0.0, 1.0) lives in GraphPPL.jl. But the @initialization macro is not part of GraphPPL, but RxInfer. The proper fix would be to move this code maybe to BayesBase.jl and let both GraphPPL and RxInfer reuse it.

[wouterwln]

Don't you think we can do a similar thing to GraphPPL where we can use the GraphPPLs default_parametrization and factor_alias in RxInfer as well to transform this code? I don't see why if RxInfer depends on GraphPPL why we wouldn't reuse this function. The actual transformation is something that happens in RxInfer anyway:

https://github.com/ReactiveBayes/RxInfer.jl/blob/e8b1215e68451baa7bfad55c541a61354eba916e/src/model/graphppl.jl#L240

[bvdmitri]

Hm, I think you're right @wouterwln