Free energy computation of Transition is broken

[wouterwln]

In the following model, we have a simple transition node:

@model function transition_test(y)
    x ~ Categorical([0.1, 0.4, 0.5])
    y ~ Transition(x, diageye(3))
end

infer(model = transition_test(), data = (y =[0, 1, 0],), free_energy=true)

and free energy is being computed. However, when we assume a joint over the x and y:

@model function transition_test(y)
    x ~ Categorical([0.1, 0.4, 0.5])
    y ~ Transition(x, diageye(3))
end

infer(model = transition_test(), data = (y = UnfactorizedData([0, 1, 0]),), free_energy=true)

The inference procedure breaks and free energy is infinite. This is because the operation (https://github.com/ReactiveBayes/ReactiveMP.jl/blob/1d28b19d3b1c96ae42dc55dba12547bd94759995/src/nodes/predefined/transition.jl#L22) here is probably not numerically stable with zeroes. Nevertheless, it is a bug.

[bvdmitri]

The comment right above the link that you've shared mentions this exact issue? Doesn't it solve it?

[bvdmitri]

Related issue from Magnus ReactiveBayes/ReactiveMP.jl#165 and PR ReactiveBayes/ReactiveMP.jl#169

[bvdmitri]

Ah, I missed UnfactorizedData in this example, yeah, indeed a problem, should be fixed