def sampler_emcee(D):
# nlinks: number of iterations for each walker
# please specify as even
D['nlinks'] = 100
# nwalkers: number of interacting chains
D['nwalkers'] = 200
# ntemps:
D['ntemps'] = 5
# ntune: number of initialization steps to discard
D['ntune'] = 500
st = time.time()
# identify starting positions for chains
tmp = start_pos(D['ntemps'] * D['nwalkers'], D)
pos0 = tmp.reshape((D['ntemps'], D['nwalkers'], D['dim']))
likelihood = D['likelihood']
# run MCMC for the model of interest
sampler = PTSampler(ntemps=D['ntemps'],
nwalkers=D['nwalkers'],
dim=len(D['pname']),
logl=likelihood,
logp=prior,
loglargs=(D, ),
logpargs=(D, ))
for pos, lnprob, lnlike in sampler.sample(pos0, iterations=D['ntune']):
pass
burntrace = sampler.chain
sampler.reset()
print(burntrace.shape)
print("burnin completed")
for pos, lnprob, lnlike in sampler.sample(pos, iterations=D['nlinks']):
pass
finaltrace = sampler.chain
print(finaltrace.shape)
D['rawtrace'] = np.concatenate((burntrace[0, ...], finaltrace[0, ...]),
axis=1)
D['sampling_time'] = np.round(time.time() - st, 2)
st = time.time()
D['lnZ'], D['dlnZ'] = sampler.thermodynamic_integration_log_evidence()
elaps = np.round(time.time() - st, 2)
msg = "model evidence evaluation: " + str(elaps) + " seconds"
WP(msg, D['wrt_file'])
return D
# The resulting samples (nsteps/thin of them) are stored as the sampler.chain # property: assert sampler.chain.shape == (ntemps, nwalkers, nsteps/10, ndim) # Chain must have shape (ntemps, nwalkers, nsteps, ndim)... # Zero temperature mean: # mu0 = np.mean(np.mean(sampler.chain[0,...], axis=0), axis=0) # Longest autocorrelation length (over any temperature) # max_acl = np.max(sampler.acor) # Compute the evidence. # API notes at http://dan.iel.fm/emcee/current/api/#the-parallel-tempered-ensemble-sampler approximation, uncertainty = sampler.thermodynamic_integration_log_evidence() # Report! print "Estimated log evidence = ",approximation,"+/-",uncertainty print " Analytic log evidence = ",log_evidence # PT burning in for 100 iterations... # PT sampling for 1000 iterations... # Estimated log evidence = -16.9586413385 +/- 7.29507626509 # Analytic log evidence = -6.67931612501 # BUG!
