def _bql_logpdf(bdb, population_id, generator_id, targets, constraints):
# P(T | C) = \sum_M P(T, M | C)
# = \sum_M P(T | C, M) P(M | C)
# = \sum_M P(T | C, M) P(M) P(C | M) / P(C)
# = \sum_M P(T | C, M) P(M) P(C | M) / \sum_M' P(C, M')
# = \sum_M P(T | C, M) P(M) P(C | M) / \sum_M' P(C | M') P(M')
#
# For a generator M, logpdf(M) computes P(T | C, M), and
# loglikelihood(M) computes P(C | M). For now, we weigh each
# generator uniformly; eventually, we ought to allow the user to
# specify a prior weight (XXX and update some kind of posterior
# weight?).
rowid, constraints = _retrieve_rowid_constraints(bdb, population_id,
constraints)
def logpdf(generator_id, metamodel):
return metamodel.logpdf_joint(bdb, generator_id, rowid, targets,
constraints, None)
def loglikelihood(generator_id, metamodel):
if not constraints:
return 0
return metamodel.logpdf_joint(bdb, generator_id, rowid, constraints,
[], None)
generator_ids = _retrieve_generator_ids(bdb, population_id, generator_id)
metamodels = [
core.bayesdb_generator_metamodel(bdb, g) for g in generator_ids
]
loglikelihoods = map(loglikelihood, generator_ids, metamodels)
logpdfs = map(logpdf, generator_ids, metamodels)
return logavgexp_weighted(loglikelihoods, logpdfs)
def _bql_logpdf(bdb, population_id, generator_id, modelnos, targets,
constraints):
# P(T | C) = \sum_M P(T, M | C)
# = \sum_M P(T | C, M) P(M | C)
# = \sum_M P(T | C, M) P(M) P(C | M) / P(C)
# = \sum_M P(T | C, M) P(M) P(C | M) / \sum_M' P(C, M')
# = \sum_M P(T | C, M) P(M) P(C | M) / \sum_M' P(C | M') P(M')
#
# For a generator M, logpdf(M) computes P(T | C, M), and
# loglikelihood(M) computes P(C | M). For now, we weigh each
# generator uniformly; eventually, we ought to allow the user to
# specify a prior weight (XXX and update some kind of posterior
# weight?).
rowid, constraints = _retrieve_rowid_constraints(
bdb, population_id, constraints)
def logpdf(generator_id, backend):
return backend.logpdf_joint(
bdb, generator_id, modelnos, rowid, targets, constraints)
def loglikelihood(generator_id, backend):
if not constraints:
return 0
return backend.logpdf_joint(
bdb, generator_id, modelnos, rowid, constraints, [])
generator_ids = _retrieve_generator_ids(bdb, population_id, generator_id)
backends = [
core.bayesdb_generator_backend(bdb, g)
for g in generator_ids
]
loglikelihoods = map(loglikelihood, generator_ids, backends)
logpdfs = map(logpdf, generator_ids, backends)
return logavgexp_weighted(loglikelihoods, logpdfs)