SeedBank Patterns

This page documents design patterns for using SeedBank.

Configurable RNGs

If you are writing code that exposes specifiable RNG seeds, such as a SciKit-style estimator or a function that uses random number generation, what we recommend doing is:

• Expose a parameter for controlling the RNG (in LensKit, we call this rng_spec).

• Pass this parameter to seedbank.numpy_rng() (or seedbank.numpy_random_state(), if you require the legacy API) to obtain a random number generator.

With this, your function can take any of the following:

• None, to use a seed dervied from the root (each time you generate a new RNG, it calls numpy.random.SeedSequence.spawn(), so you don’t use the same seed twice; this results in deterministic seeding so long as the program generates the RNGs in the same order every time).

• An integer seed.

• A string (either unicode or bytes) containing seed material.

• A numpy.random.SeedSequence

• A numpy.random.Generator or numpy.random.RandomState, in which case it is returned as-is (if the same type as the desired RNG), or its bit generator is extracted and reused in the desired type of generator.

This is equivalent to using sklearn.utils.check_random_state(), except it integrates with the SeedBank seed management infrastructure.