def _set_rejection_round(self, round):
self._update_round_info(self.state['round'])
# Get a subseed for this round for ensuring consistent results for the round
seed = self.seed if round == 0 else get_sub_seed(self.seed, round)
self._round_random_state = np.random.RandomState(seed)
self._rejection = Rejection(
self.model,
discrepancy_name=self.discrepancy_name,
output_names=self.output_names,
batch_size=self.batch_size,
seed=seed,
max_parallel_batches=self.max_parallel_batches)
def _init_new_round(self):
round = self.state['round']
dashes = '-'*16
logger.info('%s Starting round %d %s' % (dashes, round, dashes))
# Get a subseed for this round for ensuring consistent results for the round
seed = self.seed if round == 0 else get_sub_seed(self.seed, round)
self._round_random_state = np.random.RandomState(seed)
self._rejection = Rejection(self.model,
discrepancy_name=self.discrepancy_name,
output_names=self.output_names,
batch_size=self.batch_size,
seed=seed,
max_parallel_batches=self.max_parallel_batches)
self._rejection.set_objective(self.objective['n_samples'],
threshold=self.current_population_threshold)
def sample(self,
n_samples,
warmup=None,
n_chains=4,
threshold=None,
initials=None,
algorithm='nuts',
n_evidence=None,
**kwargs):
r"""Sample the posterior distribution of BOLFI.
Here the likelihood is defined through the cumulative density function
of the standard normal distribution:
L(\theta) \propto F((h-\mu(\theta)) / \sigma(\theta))
where h is the threshold, and \mu(\theta) and \sigma(\theta) are the posterior mean and
(noisy) standard deviation of the associated Gaussian process.
The sampling is performed with an MCMC sampler (the No-U-Turn Sampler, NUTS).
Parameters
----------
n_samples : int
Number of requested samples from the posterior for each chain. This includes warmup,
and note that the effective sample size is usually considerably smaller.
warmpup : int, optional
Length of warmup sequence in MCMC sampling. Defaults to n_samples//2.
n_chains : int, optional
Number of independent chains.
threshold : float, optional
The threshold (bandwidth) for posterior (give as log if log discrepancy).
initials : np.array of shape (n_chains, n_params), optional
Initial values for the sampled parameters for each chain.
Defaults to best evidence points.
algorithm : string, optional
Sampling algorithm to use. Currently only 'nuts' is supported.
n_evidence : int
If the regression model is not fitted yet, specify the amount of evidence
Returns
-------
BolfiSample
"""
if self.state['n_batches'] == 0:
self.fit(n_evidence)
# TODO: other MCMC algorithms
posterior = self.extract_posterior(threshold)
warmup = warmup or n_samples // 2
# Unless given, select the evidence points with smallest discrepancy
if initials is not None:
if np.asarray(initials).shape != (n_chains, self.target_model.input_dim):
raise ValueError("The shape of initials must be (n_chains, n_params).")
else:
inds = np.argsort(self.target_model.Y[:, 0])
initials = np.asarray(self.target_model.X[inds])
self.target_model.is_sampling = True # enables caching for default RBF kernel
tasks_ids = []
ii_initial = 0
# sampling is embarrassingly parallel, so depending on self.client this may parallelize
for ii in range(n_chains):
seed = get_sub_seed(self.seed, ii)
# discard bad initialization points
while np.isinf(posterior.logpdf(initials[ii_initial])):
ii_initial += 1
if ii_initial == len(inds):
raise ValueError(
"BOLFI.sample: Cannot find enough acceptable initialization points!")
tasks_ids.append(
self.client.apply(
mcmc.nuts,
n_samples,
initials[ii_initial],
posterior.logpdf,
posterior.gradient_logpdf,
n_adapt=warmup,
seed=seed,
**kwargs))
ii_initial += 1
# get results from completed tasks or run sampling (client-specific)
chains = []
for id in tasks_ids:
chains.append(self.client.get_result(id))
chains = np.asarray(chains)
print(
"{} chains of {} iterations acquired. Effective sample size and Rhat for each "
"parameter:".format(n_chains, n_samples))
for ii, node in enumerate(self.parameter_names):
print(node, mcmc.eff_sample_size(chains[:, :, ii]),
mcmc.gelman_rubin(chains[:, :, ii]))
self.target_model.is_sampling = False
return BolfiSample(
method_name='BOLFI',
chains=chains,
parameter_names=self.parameter_names,
warmup=warmup,
threshold=float(posterior.threshold),
n_sim=self.state['n_sim'],
seed=self.seed)
def sample(self,
n_samples,
warmup=None,
n_chains=4,
initials=None,
algorithm='nuts',
sigma_proposals=None,
n_evidence=None,
*args,
**kwargs):
"""Sample from the posterior distribution of BOLFIRE.
Sampling is performed with an MCMC sampler.
Parameters
----------
n_samples: int
Number of requested samples from the posterior for each chain. This includes warmup,
and note that the effective sample size is usually considerably smaller.
warmup: int, optional
Length of warmup sequence in MCMC sampling.
n_chains: int, optional
Number of independent chains.
initials: np.ndarray (n_chains, n_params), optional
Initial values for the sampled parameters for each chain.
algorithm: str, optional
Sampling algorithm to use.
sigma_proposals: np.ndarray
Standard deviations for Gaussian proposals of each parameter for Metropolis-Hastings.
n_evidence: int, optional
If the surrogate model is not fitted yet, specify the amount of evidence.
Returns
-------
BOLFIRESample
"""
# Fit posterior in case not done
if self.state['n_batches'] == 0:
self.fit(n_evidence)
# Check algorithm
if algorithm not in ['nuts', 'metropolis']:
raise ValueError('The given algorithm is not supported.')
# Check standard deviations of Gaussian proposals when using Metropolis-Hastings
if algorithm == 'metropolis':
if sigma_proposals is None:
raise ValueError('Gaussian proposal standard deviations have '
'to be provided for Metropolis-sampling.')
elif sigma_proposals.shape[0] != self.target_model.input_dim:
raise ValueError('The length of Gaussian proposal standard '
'deviations must be n_params.')
posterior = self.extract_result()
warmup = warmup or n_samples // 2
# Unless given, select the evidence points with best likelihood ratio
if initials is not None:
if np.asarray(initials).shape != (n_chains,
self.target_model.input_dim):
raise ValueError(
'The shape of initials must be (n_chains, n_params).')
else:
inds = np.argsort(self.target_model.Y[:, 0])
initials = np.asarray(self.target_model.X[inds])
# Enable caching for default RBF kernel
self.target_model.is_sampling = True
tasks_ids = []
ii_initial = 0
for ii in range(n_chains):
seed = get_sub_seed(self.seed, ii)
# Discard bad initialization points
while np.isinf(posterior.logpdf(initials[ii_initial])):
ii_initial += 1
if ii_initial == len(inds):
raise ValueError(
'BOLFIRE.sample: Cannot find enough acceptable '
'initialization points!')
if algorithm == 'nuts':
tasks_ids.append(
self.client.apply(mcmc.nuts,
n_samples,
initials[ii_initial],
posterior.logpdf,
posterior.gradient_logpdf,
n_adapt=warmup,
seed=seed,
**kwargs))
elif algorithm == 'metropolis':
tasks_ids.append(
self.client.apply(mcmc.metropolis,
n_samples,
initials[ii_initial],
posterior.logpdf,
sigma_proposals,
warmup,
seed=seed,
**kwargs))
ii_initial += 1
# Get results from completed tasks or run sampling (client-specific)
chains = []
for id in tasks_ids:
chains.append(self.client.get_result(id))
chains = np.asarray(chains)
logger.info(f'{n_chains} chains of {n_samples} iterations acquired. '
'Effective sample size and Rhat for each parameter:')
for ii, node in enumerate(self.parameter_names):
logger.info(f'{node} {mcmc.eff_sample_size(chains[:, :, ii])} '
f'{mcmc.gelman_rubin_statistic(chains[:, :, ii])}')
self.target_model.is_sampling = False
return BOLFIRESample(method_name='BOLFIRE',
chains=chains,
parameter_names=self.parameter_names,
warmup=warmup,
n_sim=self.state['n_sim'],
seed=self.seed,
*args,
**kwargs)
