Do you agree @bkmi ?

I like what you propose, but I haven't groked how this may affect the new store definition. I suppose we should try to merge the new store specification first then incorporate this request?

I'm still thinking about instantiating the bound outside the store... I think it's okay because the "deepest" function which will generate the bounds is most likely this one...

Perhaps we can provide an option to give a current rng, then "split" it... I'm not sure about that yet. I'm not super comfortable with these new randomness generators.

@bkmi Hi Ben, thanks for the feedback. I have adapted the current modifications to the `simplify branch as much as possible. and added some minimum tests.

Indeed we can incorporate this one later. I did not update the notebook with the bound seeding method, since the tutorial notebooks are updated in master.

the notebooks from master have been brought into simplify now.

(Thanks a lot @rogerkuou for continuing with this work which I had left halfway!)

Perhaps we can provide an option to give a current rng, then "split" it...

@bkmi do you mean to provide the possibility to set a seeded rng in Store and then propagate this to all the calls to pdf.sample()? We could avoid storing rng as an attribute into the bound, but do something like:

class Bound:
  ...
  def sample(self, n_samples, rng=None):
    rng = rng or np.random.default_rng()
    ...
  ...

In this way one would not need to seed the bound outside the store - one would seed the store instead. Also, the seeded rng in the store could be used to draw samples from the Poissonian here - which I had overlooked earlier - and which would still not be covered if we were to store rng in the bound..

This is a pull request for v0.3, which is not actively developed. Closed for now, although the topic of reproducibility is relevant for v0.4 as well.