class SMC(Sampler):
"""Sequential Monte Carlo ABC sampler."""
def __init__(self, model, discrepancy_name=None, output_names=None, **kwargs):
"""Initialize the SMC-ABC sampler.
Parameters
----------
model : ElfiModel or NodeReference
discrepancy_name : str, NodeReference, optional
Only needed if model is an ElfiModel
output_names : list, optional
Additional outputs from the model to be included in the inference result, e.g.
corresponding summaries to the acquired samples
kwargs:
See InferenceMethod
"""
model, discrepancy_name = self._resolve_model(model, discrepancy_name)
super(SMC, self).__init__(model, output_names, **kwargs)
self._prior = ModelPrior(self.model)
self.discrepancy_name = discrepancy_name
self.state['round'] = 0
self._populations = []
self._rejection = None
self._round_random_state = None
def set_objective(self, n_samples, thresholds):
"""Set the objective of the inference."""
self.objective.update(
dict(
n_samples=n_samples,
n_batches=self.max_parallel_batches,
round=len(thresholds) - 1,
thresholds=thresholds))
self._init_new_round()
def extract_result(self):
"""Extract the result from the current state.
Returns
-------
SmcSample
"""
# Extract information from the population
pop = self._extract_population()
return SmcSample(
outputs=pop.outputs,
populations=self._populations.copy() + [pop],
weights=pop.weights,
threshold=pop.threshold,
**self._extract_result_kwargs())
def update(self, batch, batch_index):
"""Update the inference state with a new batch.
Parameters
----------
batch : dict
dict with `self.outputs` as keys and the corresponding outputs for the batch
as values
batch_index : int
"""
super(SMC, self).update(batch, batch_index)
self._rejection.update(batch, batch_index)
if self._rejection.finished:
self.batches.cancel_pending()
if self.state['round'] < self.objective['round']:
self._populations.append(self._extract_population())
self.state['round'] += 1
self._init_new_round()
self._update_objective()
def prepare_new_batch(self, batch_index):
"""Prepare values for a new batch.
Parameters
----------
batch_index : int
next batch_index to be submitted
Returns
-------
batch : dict or None
Keys should match to node names in the model. These values will override any
default values or operations in those nodes.
"""
if self.state['round'] == 0:
# Use the actual prior
return
# Sample from the proposal, condition on actual prior
params = GMDistribution.rvs(*self._gm_params, size=self.batch_size,
prior_logpdf=self._prior.logpdf,
random_state=self._round_random_state)
batch = arr2d_to_batch(params, self.parameter_names)
return batch
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 _extract_population(self):
sample = self._rejection.extract_result()
# Append the sample object
sample.method_name = "Rejection within SMC-ABC"
w, cov = self._compute_weights_and_cov(sample)
sample.weights = w
sample.meta['cov'] = cov
return sample
def _compute_weights_and_cov(self, pop):
params = np.column_stack(tuple([pop.outputs[p] for p in self.parameter_names]))
if self._populations:
q_logpdf = GMDistribution.logpdf(params, *self._gm_params)
p_logpdf = self._prior.logpdf(params)
w = np.exp(p_logpdf - q_logpdf)
else:
w = np.ones(pop.n_samples)
if np.count_nonzero(w) == 0:
raise RuntimeError("All sample weights are zero. If you are using a prior "
"with a bounded support, this may be caused by specifying "
"a too small sample size.")
# New covariance
cov = 2 * np.diag(weighted_var(params, w))
if not np.all(np.isfinite(cov)):
logger.warning("Could not estimate the sample covariance. This is often "
"caused by majority of the sample weights becoming zero."
"Falling back to using unit covariance.")
cov = np.diag(np.ones(params.shape[1]))
return w, cov
def _update_objective(self):
"""Update the objective n_batches."""
n_batches = sum([pop.n_batches for pop in self._populations])
self.objective['n_batches'] = n_batches + self._rejection.objective['n_batches']
@property
def _gm_params(self):
sample = self._populations[-1]
params = sample.samples_array
return params, sample.cov, sample.weights
@property
def current_population_threshold(self):
"""Return the threshold for current population."""
return self.objective['thresholds'][self.state['round']]
def test_pdf(self, ma2):
prior = ModelPrior(ma2)
rv = prior.rvs(size=10)
assert np.allclose(prior.pdf(rv), np.exp(prior.logpdf(rv)))