Check Model

[ofSingularMind]

Hello.

It can be sometimes difficult to debug model and inference setup as-is. What is the feasibility of adding a model_check step to inference? Possibly also separately callable?

I presume we would need to check for:

• No half-edges
• No nodes with ill-defined (partially-defined) constraints
• No loops in graph not supported by initializations
• All factors (and associated variables) that are defined in the model are supported by existing nodes / update rules
• etc

Thank you 🙏

[wouterwln]

GraphPPL does check for "ill-defined" constraints (Which is the BitSetTuples error message you will get). Furthermore, I'd love to check for "No loops in graph not supported by initializations", but how do you envision this check would work? Same for "All factors (and associated variables) that are defined in the model are supported by existing nodes / update rules". Let's look at them one at a time:

• No loops in graph not supported by initializations. This one seems easy, but the factor graph you build is not the same as the dependency graph of messages/marginals. It depends on the variational constraints in a node whether we depend on an incoming message or an adjacent marginal. Therefore, in order to detect un-initialized loops, we would have to build a dependency graph for messages/marginals (which is implicitly built by ReactiveMP but actually one of the ideas for Cortex was to build this explicitly, as it is extremely difficult/impossible to extract from ReactiveMP. Then, one we have this dependency graph in place, we'd have to detect whether or not the current initialization covers the current loops.
• "All factors (and associated variables) that are defined in the model are supported by existing nodes / update rules" The problem with this is that, before running inference, we have no way of knowing what type intermediate messages will have, so we actually don't really know whether or not update rules exist, because we don't really have an idea what types of messages will be sent at all! If we would, from the graph, be able to statically determine what the type of messages will be, we solve so many more problems than just debugging, as it will make RxInfer entirely type stable and probably 50x faster or something.

Just my 2 cents on this discussion but again dev sessions are perfect moments to work on improving RxInfer in these ways.

[ofSingularMind]

Thank you for sharing 🙏