Don't allow Deterministics in the graph of observed

[ricardoV94]

Description

In #7656 we allow PyTensor deterministic operations in the graph of observed variables, but we probably don't want to allow PyMC determinstics. Doing that makes model cloning failing, for starters:

import pymc as pm

with pm.Model() as m:
    x_data = pm.Data("x_data", [0, 1, 2])
    y_data = pm.Deterministic("y", x_data + 1)
    y = pm.Normal("y", observed=y_data)

m.clone()  # ValueError: Variable name y already exists.

I am not sure about the implications of mixing the two, I suggest we be conservative and don't allow it for now, unless a good reason crops up. Also a Deterministic like this would be wastefully saved in every iteration of pm.sample!

CC @williambdean

[williambdean]

You are using the same name for the Deterministic and the likelihood. Is that intended?

For example, this works:

import pymc as pm

from pymc.model.fgraph import clone_model

with pm.Model() as m:
    x_data = pm.Data("x_data", [0, 1, 2])
    # Change
    y_data = pm.Deterministic("y_data", x_data + 1)
    y = pm.Normal("y", observed=y_data)


m_clone = clone_model(m)

[ricardoV94]

@williambdean that means my example above was not a MWE unfortunately. @cetagostini found the original bug and I don't think it was a dumb name replication because the error only happened when we cloned the model. It was a similar error message that's why I didn't careful check if my MWE was failing at the right point