def checkConstrainAVar1b(program):
ripl = get_ripl()
ripl.assume("x", "(normal 0.0 1.0)")
ripl.assume("y", "(normal 0.0 1.0)")
ripl.predict("x", label="pid")
ripl.observe("(if (tag 0 0 (flip)) x y)", 3.0)
# Not collectSamples because we depend on the trace being in a
# non-reset state at the end
collectStateSequence(ripl, "pid", infer=program)
eq_(ripl.report("pid"), 3)
def checkConstrainAVar4a(program):
# We allow downstream processing with no requests and no randomness.
ripl = get_ripl()
ripl.assume("x", "(normal 0.0 1.0)")
ripl.assume("y", "(normal 0.0 1.0)")
ripl.assume("f", "(mem (lambda () (tag 0 0 (flip))))")
ripl.predict("(if (f) (* x 5) (* y 5))", label="pid")
ripl.observe("(if (f) x y)", 3.0)
# Not collectSamples because we depend on the trace being in a
# non-reset state at the end
collectStateSequence(ripl, "pid", infer=program)
eq_(ripl.report("pid"), 15.0)
def _test_serialize_program(v, label, action):
engine = v.sivm.core_sivm.engine
if action == 'serialize':
trace1 = engine.getDistinguishedTrace()
serialized = trace1.dump()
trace2 = engine.model.restore_trace(serialized)
assert isinstance(serialized, tuple)
assert len(serialized) == 3
assert isinstance(serialized[0], list)
assert all(isinstance(x, dict) for x in serialized[0])
assert isinstance(serialized[1],
dict) # Mapping directive ids to directives
for (key, val) in serialized[1].iteritems():
assert isinstance(key, int)
assert isinstance(val, list)
assert isinstance(serialized[2], set) # Names of bound foreign sps
for elem in serialized[2]:
assert isinstance(elem, basestring)
assert isinstance(trace2, type(trace1))
assert isinstance(trace2.trace, type(trace1.trace))
elif action == 'copy':
trace1 = engine.getDistinguishedTrace()
trace2 = engine.model.copy_trace(trace1)
assert isinstance(trace2, type(trace1))
assert isinstance(trace2.trace, type(trace1.trace))
elif action == 'convert_puma':
try:
from venture.puma import trace
except ImportError:
raise SkipTest("Puma backend does not appear to be installed")
trace1 = engine.getDistinguishedTrace()
engine.to_puma()
trace2 = engine.getDistinguishedTrace()
assert 'venture.puma' in trace2.trace.__module__
elif action == 'convert_lite':
trace1 = engine.getDistinguishedTrace()
engine.to_lite()
trace2 = engine.getDistinguishedTrace()
assert 'venture.lite' in trace2.trace.__module__
else:
assert False
infer = "(resimulation_mh default one %s)" % default_num_transitions_per_sample(
)
engine.model.create_trace_pool([trace2])
r2 = collectStateSequence(v, label, infer=infer)
engine.model.create_trace_pool([trace1])
r1 = collectStateSequence(v, label, infer=infer)
return reportSameDiscrete(r1, r2)
def test_serialize_forget(seed):
with tempfile.NamedTemporaryFile(prefix='serialized.ripl') as f:
v1 = get_ripl(seed=seed)
v1.assume('is_tricky', '(flip 0.2)')
v1.assume('theta', '(if is_tricky (beta 1.0 1.0) 0.5)')
v1.assume('flip_coin', '(lambda () (flip theta))')
for i in range(10):
v1.observe('(flip_coin)', 'true', label='y{}'.format(i))
v1.infer("(incorporate)")
v1.save(f.name)
v2 = get_ripl()
v2.load(f.name)
for i in range(10):
v2.forget('y{}'.format(i))
v2.predict('is_tricky', label='pid')
infer = "(resimulation_mh default one %s)" % default_num_transitions_per_sample(
)
samples = collectStateSequence(v2, 'pid', infer=infer)
ans = [(False, 0.8), (True, 0.2)]
return reportKnownDiscrete(ans, samples)
def testHiddenDeterminism2():
# Makes sure that blocking can avoid proposing impossible things.
ripl = progHiddenDeterminism()
# TODO enumerative gibbs triggers the log(0) bug even when blocked.
predictions = collectStateSequence(ripl, "c2", infer="(mh default all 50)")
# Block MH should explore the posterior
ans = [(True, .5), (False, .5)]
return reportKnownDiscrete(ans, predictions)
def testHiddenDeterminism1():
# Makes sure that proposals of impossible things don't cause trouble
ripl = progHiddenDeterminism()
raise SkipTest("Crashes with a log(0) problem in log density of bernoulli. Issue: https://app.asana.com/0/9277419963067/10386828313646")
c1 = ripl.report("c1")
# TODO Expand collectSamples to accept a list of indices and report all of them
# TODO Expand collectSamples to accept the inference command as a string
predictions = collectStateSequence(ripl, "c1", infer="(resimulation_mh default one 20)")
# Single-site MH can't move on this problem
for pred in predictions:
eq_(pred, c1)
