def run_DREAM(self,nsamples=100000):
model = pymc.Model()
with model:
params = pymc.Normal('params', mu=self.start_parameters,
sd=np.array([1.0
]*len(self.start_parameters)),
shape=(len(self.start_parameters)))
#params = pymc.Flat('params',shape=(len(self.start_parameters)))
global cost_function
cost_function = self.cost_function
error = pymc.Potential('error', DREAM_cost(params))
nseedchains = 10*len(self.model.parameters_rules())
step = pymc.Dream(variables=[params],
nseedchains=nseedchains,
blocked=True,
start_random=False,
save_history=True,
parallel=True,
adapt_crossover=False,
verbose=False,)
trace = pymc.sample(nsamples, step,
start=self.pso_results,
njobs=self.nchains,
use_mpi=False,
progressbar=False,)
cont_flag = True
while cont_flag:
cont_flag = False
conv_stats = gelman_rubin(trace)
for i in conv_stats['params']:
if i>1.2:
print "Parameters have not converged, will continue run."
print "Value so far is %s"%i
cont_flag = True
break
trace = pymc.sample(int(nsamples*.1), step,
#start=self.pso_results,
njobs=self.nchains,
use_mpi=False,
trace = trace,
progressbar=False,)
conv_stats = gelman_rubin(trace)
for i in conv_stats['params']:
print i,i<1.2
#pymc.traceplot(trace,vars=[params,error])
#plt.show()
return trace
def TotalError(self):
"""Compute Gelman-Rubin error for a set of traces."""
assert self.traces
try:
# r_hat is the "potential scale reduction", a convergence
# diagnostic computed by comparing between-chain and
# within-chain variance.
r_hat = diagnostics.gelman_rubin(self.traces)
except FloatingPointError:
# This happens when the within-chain standard deviation is
# 0. The chain perfectly correlates with itself, implying
# complete failure to converge (if the chain has more than one
# state).
return utils.VERY_LARGE_NUMBER
return max(0, r_hat - 1)
def TotalError(self):
"""Compute Gelman-Rubin error for a set of traces."""
assert self.traces
try:
# r_hat is the "potential scale reduction", a convergence
# diagnostic computed by comparing between-chain and
# within-chain variance.
r_hat = diagnostics.gelman_rubin(self.traces)
except FloatingPointError:
# This happens when the within-chain standard deviation is
# 0. The chain perfectly correlates with itself, implying
# complete failure to converge (if the chain has more than one
# state).
return utils.VERY_LARGE_NUMBER
return max(0, r_hat - 1)
for row in zip(*uparams)
]
for name, x in zip(params_names, numpy.transpose(params_flat)):
print ' parameter %s' % name
nmin, kthin, nburn, nprec, kmind = raftery_lewis(x,
q=0.1,
r=0.01,
verbose=False)
print ' Raftery-Lewis (q=0.1, r=0.01): skip %d, use %d (of %d)' % (
nburn, nprec, len(x))
nmin, kthin, nburn, nprec, kmind = raftery_lewis(x,
q=0.025,
r=0.005,
verbose=False)
print ' Raftery-Lewis (q=0.025, r=0.005): skip %d, use %d (of %d)' % (
nburn, nprec, len(x))
results.append(params_flat)
if len(filenames) > 1:
print 'Chain group convergence'
results = numpy.transpose(results)
print results.shape
for name, xx in zip(params_names, results):
chains = numpy.transpose(xx)
print ' parameter %s:' % name, chains.shape
print ' Gelman-Rubin:', gelman_rubin(chains)
#print ' Effective sample size:', effective_n(chains)
else:
print 'Chain group convergence not computed, pass more than one dataset'
params_names += [k] * param_length uparams.append(v) params_flat = [numpy.concatenate([pi.flatten() for pi in row]) for row in zip(*uparams)] for name, x in zip(params_names, numpy.transpose(params_flat)): print ' parameter %s' % name nmin, kthin, nburn, nprec, kmind = raftery_lewis(x, q=0.1, r=0.01, verbose=False) print ' Raftery-Lewis (q=0.1, r=0.01): skip %d, use %d (of %d)' % (nburn, nprec, len(x)) nmin, kthin, nburn, nprec, kmind = raftery_lewis(x, q=0.025, r=0.005, verbose=False) print ' Raftery-Lewis (q=0.025, r=0.005): skip %d, use %d (of %d)' % (nburn, nprec, len(x)) results.append(params_flat) if len(filenames) > 1: print 'Chain group convergence' results = numpy.transpose(results) print results.shape for name, xx in zip(params_names, results): chains = numpy.transpose(xx) print ' parameter %s:' % name, chains.shape print ' Gelman-Rubin:', gelman_rubin(chains) #print ' Effective sample size:', effective_n(chains) else: print 'Chain group convergence not computed, pass more than one dataset'