def test_gsp(self):
ndags = 10
nnodes = 8
nsamples = 500
nneighbors_list = list(range(1, 3))
mean_shds = []
percent_consistent = []
for nneighbors in nneighbors_list:
print('=== nneighbors = %s ===' % nneighbors)
print("generating DAGs")
dags = directed_erdos(nnodes, nneighbors/(nnodes-1), size=ndags)
print("generating weights")
gdags = [rand_weights(dag) for dag in dags]
print("generating samples")
samples_by_dag = [gdag.sample(nsamples) for gdag in gdags]
print("computing sufficient statistics")
suffstats = [partial_correlation_suffstat(samples) for samples in samples_by_dag]
ci_testers = [MemoizedCI_Tester(partial_correlation_test, suffstat, alpha=0.1) for suffstat in suffstats]
print("running GSP")
est_dags = [
gsp(set(range(nnodes)), ci_tester, depth=4, nruns=10)
for ci_tester in ci_testers
]
# print([str(d) for d in est_dags])
shd_by_dag = np.array([est_dag.shd_skeleton(dag) for est_dag, dag in zip(est_dags, dags)])
match_by_dag = np.sum(shd_by_dag == 0)
mean_shds.append(shd_by_dag.mean())
percent_consistent.append(match_by_dag/ndags)
print("Mean SHDs:", mean_shds)
print("Percent consistent:", percent_consistent)
'fraction_correct': correct_samples / total_samples
}
return stats
if __name__ == '__main__':
from graphical_models import DAG
from graphical_models.rand import rand_weights
from conditional_independence import partial_correlation_suffstat, MemoizedCI_Tester
from graphical_model_learning import partial_correlation_test, local_gaussian_bge_score
d = DAG(arcs={(0, 1), (2, 1)})
# d = DAG(arcs={(0, 1), (0, 2), (0, 3), (0, 4), (1, 2), (1, 3), (1, 4), (2, 3), (2, 4), (3, 4)})
g = rand_weights(d)
samples = g.sample(1000)
suffstat = partial_correlation_suffstat(samples)
suffstat['samples'] = samples
initial_perm = [1, 2, 0]
ci_tester = MemoizedCI_Tester(partial_correlation_test,
suffstat,
alpha=.05)
initial_dag = permutation2dag(initial_perm, ci_tester)
# scorer = MemoizedDecomposableScore(local_gaussian_bge_score, suffstat)
scorer_bge = MemoizedDecomposableScore(local_gaussian_bge_score, suffstat)
num_meta_iterations = 30
stats_bges = [
collect_stats_mcmc(d,
initial_perm,
initial_dag,
ci_tester,
adjacent_transposition_proposer,
def compute_gaussian_interventional_suffstat(intervention_info):
return {
intervened_nodes: partial_correlation_suffstat(samples)
for intervened_nodes, samples in intervention_info.items()
}
samples = g.sample(100)
node = 5
iv_samples = g.sample_interventional({node: GaussIntervention(0, 1)}, 100)
data = {frozenset(): samples, frozenset({node}): iv_samples}
suffstat = compute_gaussian_interventional_suffstat(data)
# scorer = MemoizedDecomposableScore(local_gaussian_interventional_bic_score, suffstat)
# score = scorer.get_score(d)
for node in d.nodes:
score = local_gaussian_interventional_bic_score(
node, d.parents_of(node), suffstat, 0)
print(node, score)
print()
all_samples = np.vstack((samples, iv_samples))
suffstat = partial_correlation_suffstat(all_samples)
for node in d.nodes:
score = local_gaussian_bic_score(node, d.parents_of(node), suffstat, 0)
print(node, score)
print("***** I-BGe *****")
suffstat = compute_gaussian_interventional_suffstat(data)
# scorer = MemoizedDecomposableScore(local_gaussian_interventional_bic_score, suffstat)
# score = scorer.get_score(d)
for node in d.nodes:
score = local_gaussian_interventional_bge_score(
node, d.parents_of(node), suffstat)
print(node, score)
print()