Enhancement: Extend seeding option to a numpy RandomState/Generator

[maichmueller]

Hello,

I would like to discuss the possibility of adding the option of passing a numpy RandomState or Generator as 'seed', in order to seed via a client maintained rng state machine.

SciPy allows the seed to be used like this:

seed{None, int, RandomState, Generator}, optional
This parameter defines the object to use for drawing random variates. If seed is None the RandomState singleton is used. If seed is an int, a new RandomState instance is used, seeded with seed. If seed is already a RandomState or Generator instance, then that object is used. Default is None.

as is mentioned here.

This would have the following advantages:

• Thread-safe seeding
• Non-interference by other code (setting the seed via np.random.seed(seed) is vulnerable to outside manipulation and prone to create hard to detect bugs)
• Client decided RNG source if needed, but especially client maintained and manipulated if desired.

I am unfamiliar with the pymc3 code base and have only seen the 'seed' parameter used via np.random.seed in the few files I have searched, but it seemed like it could be generally made to support generators.

What are your thoughts on this?

Cheers,
Michael