def _joint_logpdf(self, bdb, genid, modelno, Q, Y, n_samples=None):
# XXX Computes the joint probability of query Q given evidence Y
# for a single model. The function is a likelihood weighted
# integrator.
# XXX Determine.
if n_samples is None:
n_samples = self.n_samples
# Validate inputs.
if modelno is None:
raise ValueError('Invalid modelno None, integer requried.')
if len(Q) == 0:
raise ValueError('Invalid query Q: len(Q) == 0.')
# Ensure consistency of any duplicates in Q and Y.
Q = self._queries_consistent_with_constraints(Q, Y)
if Q is None:
return float('-inf')
for r, _, _ in Q+Y:
assert r == Q[0][0], "Cannot assess more than one row, "\
"%s and %s requested" % (Q[0][0], r)
# (Q,Y) marginal joint density.
_, QY_weights = self._weighted_sample(bdb, genid, modelno,
Q[0][0], Q+Y, n_samples=n_samples)
# Y marginal density.
_, Y_weights = self._weighted_sample(bdb, genid, modelno,
Q[0][0], Y, n_samples=n_samples)
# XXX TODO Keep sampling until logpQY <= logpY
logpQY = logmeanexp(QY_weights)
logpY = logmeanexp(Y_weights)
return logpQY - logpY
def bql_row_column_predictive_probability(bdb, population_id, generator_id,
rowid, colno):
value = core.bayesdb_population_cell_value(bdb, population_id, rowid,
colno)
if value is None:
return None
# Retrieve all other values in the row.
row_values = core.bayesdb_population_row_values(bdb, population_id, rowid)
variable_numbers = core.bayesdb_variable_numbers(bdb, population_id, None)
# Build the constraints and query from rowid, using a fresh rowid.
fresh_rowid = core.bayesdb_population_fresh_row_id(bdb, population_id)
query = [(colno, value)]
constraints = [(col, value)
for (col, value) in zip(variable_numbers, row_values)
if (value is not None) and (col != colno)]
def generator_predprob(generator_id):
metamodel = core.bayesdb_generator_metamodel(bdb, generator_id)
return metamodel.logpdf_joint(bdb, generator_id, fresh_rowid, query,
constraints, None)
generator_ids = _retrieve_generator_ids(bdb, population_id, generator_id)
predprobs = map(generator_predprob, generator_ids)
r = logmeanexp(predprobs)
return ieee_exp(r)
def bql_row_column_predictive_probability(bdb, population_id, generator_id,
modelnos, rowid, targets,
constraints):
targets = json.loads(targets)
constraints = json.loads(constraints)
modelnos = _retrieve_modelnos(modelnos)
# Build the constraints and query from rowid, using a fresh rowid.
fresh_rowid = core.bayesdb_population_fresh_row_id(bdb, population_id)
def retrieve_values(colnos):
values = [
core.bayesdb_population_cell_value(bdb, population_id, rowid,
colno) for colno in colnos
]
return [(c, v) for (c, v) in zip(colnos, values) if v is not None]
cgpm_targets = retrieve_values(targets)
# If all targets have NULL values, return None.
if len(cgpm_targets) == 0:
return None
cgpm_constraints = retrieve_values(constraints)
def generator_predprob(generator_id):
metamodel = core.bayesdb_generator_metamodel(bdb, generator_id)
return metamodel.logpdf_joint(bdb, generator_id, modelnos, fresh_rowid,
cgpm_targets, cgpm_constraints)
generator_ids = _retrieve_generator_ids(bdb, population_id, generator_id)
predprobs = map(generator_predprob, generator_ids)
r = logmeanexp(predprobs)
return ieee_exp(r)
def logpdf_joint(
self,
bdb,
generator_id,
modelnos,
rowid,
targets,
_constraints,
):
# Note: The constraints are irrelevant for the same reason as
# in simulate_joint.
(all_mus, all_sigmas) = self._all_mus_sigmas(bdb, generator_id)
def model_log_pdf(modelno):
mus = all_mus[modelno]
sigmas = all_sigmas[modelno]
def logpdf_1((colno, x)):
return self._logpdf_1(bdb, generator_id, mus, sigmas, colno, x)
return sum(map(logpdf_1, targets))
# XXX Ignore modelnos and aggregate over all of them.
modelwise = [model_log_pdf(m) for m in sorted(all_mus.keys())]
return logmeanexp(modelwise)
def bql_row_column_predictive_probability(
bdb, population_id, generator_id, modelnos, rowid, targets,
constraints):
targets = json.loads(targets)
constraints = json.loads(constraints)
modelnos = _retrieve_modelnos(modelnos)
# Build the constraints and query from rowid, using a fresh rowid.
fresh_rowid = core.bayesdb_population_fresh_row_id(bdb, population_id)
def retrieve_values(colnos):
values = [
core.bayesdb_population_cell_value(bdb, population_id, rowid, colno)
for colno in colnos
]
return [(c,v) for (c,v) in zip (colnos, values) if v is not None]
cgpm_targets = retrieve_values(targets)
# If all targets have NULL values, return None.
if len(cgpm_targets) == 0:
return None
cgpm_constraints = retrieve_values(constraints)
def generator_predprob(generator_id):
backend = core.bayesdb_generator_backend(bdb, generator_id)
return backend.logpdf_joint(
bdb, generator_id, modelnos, fresh_rowid, cgpm_targets,
cgpm_constraints)
generator_ids = _retrieve_generator_ids(bdb, population_id, generator_id)
predprobs = map(generator_predprob, generator_ids)
r = logmeanexp(predprobs)
return ieee_exp(r)
def logpdf_joint(self, bdb, generator_id, targets, constraints, modelno):
if modelno is None:
modelnos = core.bayesdb_generator_modelnos(bdb, generator_id)
else:
modelnos = [modelno]
with bdb.savepoint():
return logmeanexp([self._joint_logpdf(bdb, generator_id, modelno,
targets, constraints) for modelno in modelnos])
def logpdf_joint(self, bdb, generator_id, targets, _constraints,
modelno=None):
# Note: The constraints are irrelevant for the same reason as
# in simulate_joint.
(all_mus, all_sigmas) = self._all_mus_sigmas(bdb, generator_id)
def model_log_pdf(modelno):
return sum(logpdf_gaussian(value, all_mus[modelno][colno],
all_sigmas[modelno][colno])
for (_, colno, value) in targets)
modelwise = [model_log_pdf(m) for m in sorted(all_mus.keys())]
return logmeanexp(modelwise)
def logpdf_joint(self, bdb, generator_id, modelnos, rowid, targets,
_constraints,):
# Note: The constraints are irrelevant for the same reason as
# in simulate_joint.
(all_mus, all_sigmas) = self._all_mus_sigmas(bdb, generator_id)
def model_log_pdf(modelno):
mus = all_mus[modelno]
sigmas = all_sigmas[modelno]
def logpdf_1((colno, x)):
return self._logpdf_1(bdb, generator_id, mus, sigmas, colno, x)
return sum(map(logpdf_1, targets))
# XXX Ignore modelnos and aggregate over all of them.
modelwise = [model_log_pdf(m) for m in sorted(all_mus.keys())]
return logmeanexp(modelwise)
def bql_row_column_predictive_probability(bdb, population_id, generator_id,
rowid, colno):
value = core.bayesdb_population_cell_value(
bdb, population_id, rowid, colno)
if value is None:
return None
def generator_predprob(generator_id):
metamodel = core.bayesdb_generator_metamodel(bdb, generator_id)
return metamodel.logpdf_joint(
bdb, generator_id, [(rowid, colno, value)], [], None)
generator_ids = [generator_id] if generator_id is not None else \
core.bayesdb_population_generators(bdb, population_id)
predprobs = map(generator_predprob, generator_ids)
r = logmeanexp(predprobs)
return ieee_exp(r)
def test_logmeanexp():
assert logmeanexp([-1000., -1000.]) == -1000.
assert relerr(math.log(0.5 * (1 + math.exp(-1.))), logmeanexp([0., -1.])) \
< 1e-15
assert relerr(math.log(0.5), logmeanexp([0., -1000.])) < 1e-15
def test_logmeanexp():
assert logmeanexp([-1000., -1000.]) == -1000.
assert relerr(math.log(0.5 * (1 + math.exp(-1.))), logmeanexp([0., -1.])) \
< 1e-15
assert relerr(math.log(0.5), logmeanexp([0., -1000.])) < 1e-15
