How-to Guides

How to save and resume long computation

RandomState is pickleable. Pickling allows to save and restore the internal state of the pseudo-random number generators.

Saving state of pseudo-random basic random number generators

    import numpy as np
    import mkl_random
    import pickle

    rs = mkl_random.RandomState(seed=777, brng="r250")
    draw = rs.standard_normal(size=1357913)

    # pickle random state
    saved = pickle.dumps(rs)

    # draw some numbers as if computation were to continue
    post_draw = rs.gamma(5, 1, size=100)

    # restore random state, and continue from
    restored_rs = pickle.loads(saved)
    resumed_draw = restored_rs.gamma(5, 1, size=100)

    # sample from restored stated is the same as sample
    # from the original one
    assert np.array_equal(restored_rs, resumed_draw)

Stochastic computations in parallel with multiprocessing

When performing stochastic computations in parallel, care is due to ensure statistical independence of samples drawn in parallel.

Basic quasi-random number generators provide different means to accomplishing this. Some support skipahead() method or leapfrog() method, while others provide a fixed-size family of generators with nice property that generators from such family, initialized equally, produce streams of randomness statistically indistinguishable from independent.

skipahead(nskips)

Advance the state of the generator using skip-ahead method, or raise ValueError exception if not supported.

The argument nskips must be a positive Python integer.

The method is supported for “philox4x32x10”, “mrg32k3a”, “mcg31m1”, “mcg59”, “wh”, “mt19937”, “sfmt19937”, and “ars5” basic random number generators.

Note

When using skipahead(), it is important to ensure that a parallel task does not consume more than nskips states, otherwise streams of randomness begin to overlap and the assumption of statistical independence breaks down.

leapfrog(k, nstreams)

Initialize the state of the generator using leap-frog method, or raise ValueError exception if not supported.

The leap-frog method partitions state tragectory into nstream interleaved non-overlapping sub-sequences, and argument k identifies the subsequence.

The method is supported for “mcg31m1”, “mcg59”, and “wh” basic pseudo-random number generators.

Note

When using leapfrog() or skipahead() methods one must remember that parallel tasks partition generators period and choose a generator with sufficiently long period to avoid cycling over the period more than once, as doing so also breaks the assumption of statistical independence and may compromise correctness of the simulation.

mkl_random also provides two families of basic pseudo-random number generators, “mt2203” and “wh”, with property that members from particular family, initialized equally, produce streams of randomness stasistically indistunguishable from independent. To use such families in parallel computation, assign difference family generators to different parallel workers and sample those assigned generators in each parallel worker. Please refer to “examples/” folder in the GitHub repo for more details.