Seed argument for functions with RNGs about bctpy HOT 8 CLOSED

from bctpy.

it's probably not desirable for most users if bctpy behaves deterministically by default

Agreed. I was thinking something like

import numpy as np

def randmio_dir(adj, itr, seed=None):
    rng = np.random if seed is None else np.random.RandomState(seed)
    # the rest of the function, using rng

The path of least resistance may be to use the patch strategy above in a decorator:

from unittest import mock
from functools import wraps

def seedable(fn):
    @wraps(fn)
    def wrapped(*args, seed=None, **kwargs):
        if seed is None:
            return fn(*args, **kwargs)
        else:
            with mock.patch('numpy.random', np.random.RandomState(seed)):
                return fn(*args, **kwargs)

@seedable
def randmio_dir(adj, itr):
    """decorated version now has a ``seed`` kwarg"""
    # the rest of the function with no changes

It's worth noting that (annoyingly) RandomState doesn't have exactly the same functions available as np.random - it's notably missing the random/ranf/sample aliases for random_sample.

from bctpy.

from bctpy.

Not a problem! Just wanted to make sure it fit with your design goals before starting on it myself. For my own documentation, another approach to the decorator would be:

from functools import wraps

def seedable(fn):
    @wraps(fn)
    def wrapped(*args, seed=None, **kwargs):
        if seed is None:
            return fn(*args, **kwargs)
        else:
            state = np.random.get_state()
            np.random.seed()
            fn(*args, **kwargs)
            np.random.set_state(state)

Feels less hacky because we're not patching over stuff, but would probably break in threaded contexts. Then again, so might the patch.

from bctpy.

Confirmed, the patch approach also breaks threaded usage. Looks like the decorator is a no-go. Probably better for readability anyway.

from bctpy.

I think I've caught all of the cases where one randomising bct function calls another randomising bct function, and pass a deterministic seed through, but if any of them are calling out to other functions which use np.random under the hood I'm unlikely to have found them.

from bctpy.

Fixed in #70

from bctpy.

Thanks!

Since this was initially raised, numpy updates have deprecated this RandomState API in favour of the Generator API. Fortunately the design (i.e. the get_rng function) should make it easy to drop in the new RNG in without much disruption. I'll raise another issue for that though.

from bctpy.