Graphviz does not properly show observed ConstantData when subjected to intermediate Ops

[larryshamalama]

with pm.Model() as model:
    data = pm.ConstantData("data", [0, 0, 1, 1], dtype=np.int64)
    p = pm.Beta("p", 1, 1)
    bern = pm.Bernoulli("response", p, observed=data)
    
pm.model_to_graphviz(model)

yields

In other words, there should be an arrow from response to data. This is because model["data"].tag.observations is in fact an Elemwise operation rather than data as defined under the context manager. I don't think that this was broken from the recently merged PR #5634.

[ricardoV94]

I don't know the internals of model_graph, but one thing worth noting is that data only exists in the logp graph (as do other elements like potentials).

[larryshamalama]

On my end, I'm not as familiar with the internals of the initialization of the logp graph under a pm.Model(), but the observed nodes in graphviz are obtained by checking observed_rv.tag.observations as per this line.

[drbenvincent]

not 100% sure, but this is cropping up quite a bit. Eg

(from here https://www.pymc.io/projects/examples/en/latest/causal_inference/excess_deaths.html)

Personally, I think this is a relatively high priority defect

[larryshamalama]

Can try to have a closer look before the weekend. Thanks for sharing!

[larryshamalama]

not 100% sure, but this is cropping up quite a bit. Eg (from here https://www.pymc.io/projects/examples/en/latest/causal_inference/excess_deaths.html)

Personally, I think this is a relatively high priority defect

I will submit a PR that would close this issue, but I am doubtful that it would fix this problem (which should be very fixable). I will investigate further soon.