How can emcee be run deterministically?

I want to be able to run emcee multiple times and get exactly the same answer. How can this be accomplished? It does not seem to be sufficient to use random.seed(0), or numpy.random.seed(0), or to capture numpy.random.mtrand.RandomState(0) and assign it to the sampler, or in fact all three.

This is relevant, among others, for making sure that our example notebook doesn’t change when rerun, so that it can be checked in to version control without endless merge conflicts.

A quick test in raw emcee; this passes, mystifyingly, sometimes but not always:

def test_raw_emcee():
    r = []
    for i in range(2):
        random.seed(0)
        numpy.random.seed(0)
        s = numpy.random.mtrand.RandomState(0)

        def log_prob(x, ivar):
            return -0.5 * np.sum(ivar * x ** 2)

        ndim, nwalkers = 5, 100
        ivar = 1. / np.random.rand(ndim)
        # r.append(ivar[0])
        p0 = np.random.randn(nwalkers, ndim)
        #r.append(p0[0, 0])

        sampler = emcee.EnsembleSampler(nwalkers, ndim, log_prob, args=[ivar])
        sampler.random_state = s
        sampler.run_mcmc(p0, 100)

        samples = sampler.chain.reshape((-1, ndim))
        r.append(samples[0, 0])
    assert r[0] == r[1]

[aarchiba]

Thank you, that works great! In the event I think we’re going to depend on a recent version of emcee in order to have long double walkers.

[Alexis-Prel]

@aarchiba would you mind letting me know if #324 indeed addresses your needs?