def run_hmc_sampler(model, sps, lnprobf, initial_center, rp, pool=None):
"""Run a (single) HMC chain, performing initial steps to adjust the
epsilon.
"""
import hmc
sampler = hmc.BasicHMC()
eps = 0.
# need to fix this:
length = niter = nsegmax = nchains = None
# initial conditions and calulate initial epsilons
pos, prob, thiseps = sampler.sample(initial_center,
model,
iterations=10,
epsilon=None,
length=length,
store_trajectories=False,
nadapt=0)
eps = thiseps
# Adaptation of stepsize
for k in range(nsegmax):
# Advance each sampler after adjusting step size
afrac = sampler.accepted.sum() * 1.0 / sampler.chain.shape[0]
if afrac >= 0.9:
shrink = 2.0
elif afrac <= 0.6:
shrink = 1 / 2.0
else:
shrink = 1.00
eps *= shrink
pos, prob, thiseps = sampler.sample(sampler.chain[-1, :],
model,
iterations=iterations,
epsilon=eps,
length=length,
store_trajectories=False,
nadapt=0)
# this should not have actually changed during the sampling
alleps[k] = thiseps
# Main run
afrac = sampler.accepted.sum() * 1.0 / sampler.chain.shape[0]
if afrac < 0.6:
eps = eps / 1.5
hpos, hprob, eps = sampler.sample(initial_center,
model,
iterations=niter,
epsilon=eps,
length=length,
store_trajectories=False,
nadapt=0)
return sampler
def run_hmc(initial,
model,
length=100,
nsegmax=20,
iterations=5000,
adapt_iterations=100,
verbose=False):
"""Run an HMC sampler for the given number of steps after a
certain amount of adaptation. returns the sampler object.
"""
hsampler = hmc.BasicHMC(verbose=verbose)
pos, prob, teps = hsampler.sample(initial,
model,
iterations=10,
epsilon=None,
length=length,
store_trajectories=False,
nadapt=0)
eps = teps
# Adapt epsilon while burning-in
for k in range(nsegmax):
#advance each sampler after adjusting step size
afrac = hsampler.accepted.sum() * 1.0 / hsampler.chain.shape[0]
if afrac >= 0.9:
shrink = 2.0
elif afrac <= 0.6:
shrink = 1 / 2.0
else:
shrink = 1.00
eps *= shrink
pos, prob, teps = hsampler.sample(hsampler.chain[-1, :],
model,
epsilon=eps,
length=length,
iterations=adapt_iterations,
store_trajectories=False,
nadapt=0)
# Production
hsampler.sample(hsampler.chain[-1, :],
model,
iterations=iterations,
length=length,
epsilon=eps,
store_trajectories=True)
return hsampler
raxes[i, 0].imshow(stamp.pixel_values.T, origin='lower')
raxes[i, 1].imshow(im.T, origin='lower')
resid = raxes[i, 2].imshow((stamp.pixel_values - im).T,
origin='lower')
rfig.colorbar(resid, ax=raxes[i, :].tolist())
labels = ['Data', 'Model', 'Data-Model']
[ax.set_title(labels[i]) for i, ax in enumerate(raxes[0, :])]
# --- Sampling -----
if False:
p0 = ptrue.copy() #model.renormalize_theta(ptrue)
import hmc
#initialize sampler and sample
sampler = hmc.BasicHMC(model, verbose=False)
eps = sampler.find_reasonable_epsilon(p0)
#eps = 0.01
#sys.exit()
iterations = 50
length = 200
pos, prob, eps = sampler.sample(p0 * 1.1,
iterations=iterations,
epsilon=eps,
length=length,
store_trajectories=True)
vals = pos # model.renormalize_theta(pos)
rfig, raxes = pl.subplots(len(stamps), 3, sharex=True, sharey=True)
raxes = np.atleast_2d(raxes)
for i, stamp in enumerate(stamps):