model_inference.py

'''
Created on Jun 1, 2017

@author: zwieback
'''
import pymc3 as pm
import numpy as np
from pymc3.backends.base import merge_traces
def model_inference(model, niter=2000, nadvi=200000, ntraceadvi=1000, seed=123, nchains=2):
    with model:
        v_params = pm.variational.advi(n=nadvi, random_seed=seed)
        tracevi = pm.variational.sample_vp(v_params, draws=ntraceadvi, random_seed=seed)
        traces = []
        for chain in range(nchains):
            step = pm.NUTS(scaling=np.power(model.dict_to_array(v_params.stds), 2), is_cov=True, target_accept=0.95)
            trace = pm.sample(niter, chain=chain, step=step, random_seed=seed)
            trace = trace[niter//2::2]
            traces.append(trace)
        trace = merge_traces(traces)
    return trace, v_params, tracevi