alex-schaaf / ABCsampler

Simple model class and non-vectorized samplers for Approximate Bayesian Statistics (ABC)

ABCsampler

Simple model class and non-vectorized samplers for Approximate Bayesian Statistics (ABC) for use with slow simulator functions where sampling speed doesn't really matter.

How to use

Model construction

from abcsamplers import ABCModel

def simulator(prior_sample):
    return list(prior_sample.values())


def summary(simulator_out):
    return np.sum(simulator_out)


def distance(s1, observed):
    return np.abs(observed - s1)


groups = ["Prior 1", "Prior 2", "Prior 3", "Prior 4"]
priors = {group:scipy.stats.norm(0, 1) for group in groups}

model = abcsamplers.ABCModel(priors, simulator, summary, distance

Rejection sampling

from abcsamplers import REJ

observed = [4]
rej = abcsamplers.REJ(model, observed)
rej.sample(1000, threshold=0.1)

Sequential Monte Carlo sampling

from abcsamplers import SMC

observed = [4]
smc = abcsamplers.SMC(model, observed)
thresholds = [0.5, 0.4, 0.3, 0.2, 0.1, 0.05]
smc.sample(1000, thresholds)

Features

• Models wih simple hierarchy like priors -> simulator -> summary -> distance
• Graph-based model construction allowing for complex prior relationships, multiple summary statistics and distance functions

Dependencies

pip install tqdm scipy