How should I debug loops that give the variables have not been updated error?

[meditans]

Hi all, I read the tutorial on model initialization, but in practice I often find myself confused about the specifics. In the first version of this question, I included some concrete example, but I was later able to debug that myself. What through me off is that adding a MeanField() constraint to the inference of a model can trigger the warning about non initialized variables even in the case the inference would have worked without MeanField.

I wanted to leave this issue here because I wonder what could be done to help tracing better what exactly is the problem when the Variables ... have not been updated error pops up. In my experience, the message just mentions almost all the variables in the model. What prevents the automatic detection of the loop?
It also would be nice to highlight what's the difference between specifying the marginals or the message, when one could use one or the other, and how to determine what good values are.

[bvdmitri]

We indeed need a better way to debug it. For now the best shot is to try to visualize the structure of the model and backtrack the loops in the model. Its more subtle with automatic detection of the loop, because its not only about the loop in the graph structure, but instead the loop in the dependency structure. E.g. a loop in the graph structure can be broken up with MeanField constraints. With structured-MeanField the situations is more complex. There should be enough information in RxInfer to figure this out, but we don't have code that is doing precisely that