Dynesty Sampler

[Felipe-4]

This is not a bug, or a problem, but I didn't find any specific bilby forum.

It seems to me that there are differences between the dynamic nested sampling implemented by dynesty and the one by bilby. I want to use the bilby one, but I need a sampler with the same syntax of the dynesty software, that is, I just pass the loglikelihood and the prior transform function, from the unit cube to my actual prior. Is there such a function in bilby?

[ColmTalbot]

Hi @Felipe-4!

This is not the neatest thing to do as the changes are pretty strongly integrated into Bilby, but it can be done.

I've put in a snippet that works for me combing this dynesty test with logic from this method. This specifically uses the act-walk method, but with minimal tinkering, it should be able to get the acceptance-walk method to work.

import numpy as np
import dynesty
from bilby.core.sampler.base_sampler import _SamplingContainer
from bilby.core.sampler.dynesty_utils import LivePointSampler, ACTTrackingRWalk

nlive = 1000
# printing = get_printing()

# EGGBOX


# see 1306.2144
def loglike_egg(x):
    logl = ((2 + np.cos(x[0] / 2) * np.cos(x[1] / 2))**5)
    return logl


def prior_transform_egg(x):
    return x * 10 * np.pi


LOGZ_TRUTH = 235.856


def test_bounds(bound, sample):
    # stress test various boundaries
    ndim = 2
    # rstate = get_rstate()
    sampler = dynesty.NestedSampler(loglike_egg,
                                    prior_transform_egg,
                                    ndim,
                                    nlive=nlive,
                                    bound=bound,
                                    sample=sample,
                                    walks=10,  # note there is an additional needed waveform kwargs
                                    )
    sampler.run_nested(dlogz=0.01, print_progress=True)
    assert (abs(LOGZ_TRUTH - sampler.results.logz[-1]) <
            5. * sampler.results.logzerr[-1])


# here I register the new sampling methods
dynesty.dynamicsampler._SAMPLERS["live"] = LivePointSampler
dynesty.nestedsamplers._SAMPLING["act-walk"] = ACTTrackingRWalk()
# this is an unfortunately needed Bilby argument
_SamplingContainer.maxmcmc = 100000

test_bounds(bound="live", sample="act-walk")

I'm going to move this to a discussion rather than an issue, as I think that is more appropriate for this.