Random number generation

gpyreg.rng

The functions of gpyreg that draw random numbers (GP.fit, GP.random_function, SliceSampler, f_min_fill) take an rng argument resolved by the helpers below. rng=None (the default) keeps NumPy’s global legacy stream, as before generators were supported. The legacy stream uses a stateless proxy, so copying or pickling a sampler does not capture NumPy’s global state; np.random.seed continues to control its draws. Samplers pickled before the rng argument was added also resume using the global stream. A sampler with an explicit generator preserves that generator’s state when copied or pickled.

Resolving a generator or an already resolved legacy proxy returns the same object. GP.fit resolves its argument once and shares the stream between the initial design and the sampler, including when given an integer seed.

resolve_rng

gpyreg.rng.resolve_rng(rng=None)

Return the object to draw random numbers from.

Parameters:

rngNone, numpy.random.Generator, int, array_like[int], SeedSequence or BitGenerator, optional

None (the default everywhere in gpyreg) returns a picklable, stateless proxy for NumPy’s global legacy stream, forwarding draws call for call. Copying or pickling the proxy does not capture the global random state. A numpy.random.Generator or an already resolved proxy is returned as is, so streams can be shared with the caller. Anything else is passed to numpy.random.default_rng and seeds a new generator.

Returns:

rngnumpy.random.Generator or legacy stream proxy

Both expose random(), uniform(size=...), standard_normal(size) and shuffle(x) with the same meaning; see random_integer() for the one method whose name differs.

random_integer

gpyreg.rng.random_integer(rng, high)

One integer drawn uniformly from 0, ..., high - 1 (randint on the legacy proxy, integers on a Generator).