def sample_interventional_perfect(self, interventions: PerfectIntervention, nsamples: int = 1) -> np.array:
"""
Return `nsamples` samples from the graph under a perfect intervention
Parameters
----------
interventions:
nsamples:
Returns
-------
(nsamples x nnodes) matrix of samples.
"""
samples = np.zeros((nsamples, len(self._nodes)))
noise = np.zeros((nsamples, len(self._nodes)))
for ix, (node, mean, var) in enumerate(zip(self._node_list, self._means, self._variances)):
interventional_dist = interventions.get(node)
if interventional_dist is not None:
noise[:, ix] = interventional_dist.sample(nsamples)
else:
noise[:, ix] = np.random.normal(loc=mean, scale=var ** .5, size=nsamples)
t = self.topological_sort()
for node in t:
ix = self._node2ix[node]
parents = self._parents[node]
if node not in interventions and len(parents) != 0:
parent_ixs = [self._node2ix[p] for p in self._parents[node]]
parent_vals = samples[:, parent_ixs]
samples[:, ix] = np.sum(parent_vals * self._weight_mat[parent_ixs, node], axis=1) + noise[:, ix]
else:
samples[:, ix] = noise[:, ix]
return samples
def logpdf(self, samples: np.array, interventions: PerfectIntervention = None, exclude_intervention_prob=True) -> np.array:
# TODO this is about 10x slower than using multivariate_normal.logpdf with the covariance matrix
# TODO can I speed this up? where is the time spent?
sorted_nodes = self.topological_sort()
nsamples = samples.shape[0]
log_probs = np.zeros(nsamples)
if interventions is None:
for node in sorted_nodes:
node_ix = self._node2ix[node]
parent_ixs = [self._node2ix[p] for p in self._parents[node]]
if len(parent_ixs) != 0:
parent_vals = samples[:, parent_ixs]
correction = (parent_vals * self._weight_mat[parent_ixs, node]).sum(axis=1)
else:
correction = 0
log_probs += norm.logpdf(samples[:, node_ix] - correction, scale=self._variances[node_ix]**.5)
else:
for node in sorted_nodes:
node_ix = self._node2ix[node]
iv = interventions.get(node)
if iv is not None:
if not exclude_intervention_prob:
if isinstance(iv, GaussIntervention):
log_probs += iv.logpdf(samples[:, node_ix])
else:
log_probs += np.log(iv.pdf(samples[:, node_ix]))
else:
parent_ixs = [self._node2ix[p] for p in self._parents[node]]
parent_vals = samples[:, parent_ixs]
correction = (parent_vals * self._weight_mat[parent_ixs, node]).sum(axis=1)
log_probs += norm.logpdf(samples[:, node_ix] - correction, scale=self._variances[node_ix]**.5)
return log_probs