def observe(self, expression, value, label=None, type=False):
if label is None:
i = {'instruction':'observe', 'expression':expression,
'value':value}
else:
label = _symbolize(label)
i = {'instruction':'labeled_observe', 'expression':expression,
'value':value, 'label':label}
weights = self.execute_instruction(i)['value']
return v.vector(weights) if type else weights
def forget(self, label_or_did, type=False):
(tp, val) = _interp_label_or_did(label_or_did)
if tp == 'did':
i = {'instruction': 'forget', 'directive_id': val}
# if asked to forget prelude instruction, decrement _n_prelude
if label_or_did <= self._n_prelude:
self._n_prelude -= 1
else:
# assume that prelude instructions don't have labels
i = {'instruction': 'labeled_forget', 'label': val}
weights = self.execute_instruction(i)['value']
return v.vector(weights) if type else weights
def particulate(num_obs, epsilon):
ripl = get_ripl()
ripl.load_prelude()
ripl.assume("mu", "(multivariate_normal (zeros 2) (id_matrix 2))")
# A slow procedure to compute f(m) = m[0:2] * 1.0
ripl.assume(
"f",
"(lambda (m) (map (lambda (i) (* 1.0 (lookup m i))) (range 0 2)))")
ripl.assume("y",
"(lambda () (multivariate_normal (f mu) (id_matrix 2)))")
for _ in range(num_obs):
ripl.observe("(y)", val.vector(scipy.stats.norm.rvs(0, 1.0, 2)))
ripl.infer("(mh default all 1)")
def do_infer():
ripl.infer("(subsampled_mh default all 10 3 %f false 0 false 10)" %
epsilon)
return do_infer
def testVector():
# Test that array-like objects don't get interpreted as expressions.
ripl = get_ripl()
ripl.predict(v.vector(numpy.array([1, 2])))
def force(self, expression, value, type=False):
i = {'instruction': 'force', 'expression': expression, 'value': value}
weights = self.execute_instruction(i)["value"]
return v.vector(weights) if type else weights
