Create a RandomState Variable constructor Op

[brandonwillard]

We need an Op that will construct RandomState Variables with the seed values as inputs.

We should need only one such Op that handles all RandomState types (e.g. by having a fixed Op-level property that determines which types the Op generates).

This—combined with #738—will largely remove the need to use shared variables for RNG states.

[ricardoV94]

I like that approach. It sounds a lot like what we ended up doing here manually with shared RandomState variables: https://github.com/pymc-devs/pymc/blob/b29124bc054c47ca2de99789d9894db0caa17c13/pymc/initial_point.py#L200-L219

[brandonwillard]

I like that approach. It sounds a lot like what we ended up doing here manually with shared RandomState variables: https://github.com/pymc-devs/pymc/blob/b29124bc054c47ca2de99789d9894db0caa17c13/pymc/initial_point.py#L200-L219

Another way to put it: we need an Op for numpy.random.default_rng.

[kc611]

Took a stab at this and ran into an obvious problem. How do we return a symbolic representation of a SharedVariable (RandomStateSharedVariables in this case) in the Op.make_node without providing it a value in Op.make_node itself ?

[brandonwillard]

Took a stab at this and ran into an obvious problem. How do we return a symbolic representation of a SharedVariable (RandomStateSharedVariables in this case) in the Op.make_node without providing it a value in Op.make_node itself ?

There are no shared variables in this case. That's why we want this Op.

[kc611]

Do we return Variable that can contain/return RandomState symbolically?

I am having trouble understanding how such an Op will be made with current Variable classes.

[brandonwillard]

Yeah, a plain Variable can be used, but it might be better to have a specific variable type. That would make it possible to do type checking on the variable instances instead of their .type class attributes.