class ReducedTraces(Engine):
def __init__(self):
self.engine_type = 'reduced traces'
self.assumes = {} # id -> evalnode
self.observes = {} # id -> evalnode
self.predicts = {} # id -> evalnode
self.directives = []
self.db = RandomChoiceDict()
self.weighted_db = WeightedRandomChoiceDict()
self.choices = {} # hash -> evalnode
self.xrps = {} # hash -> (xrp, set of application nodes)
self.env = EnvironmentNode()
self.p = 0
self.uneval_p = 0
self.eval_p = 0
self.new_to_old_q = 0
self.old_to_new_q = 0
self.debug = False
# necessary because of the new XRP interface requiring some state kept while doing inference
self.application_reflip = False
self.reflip_node = ReducedEvalNode(self, self.env, VarExpression(''))
self.nodes = []
self.old_vals = [Value()]
self.new_vals = [Value()]
self.old_val = Value()
self.new_val = Value()
self.reflip_xrp = XRP()
self.mhstats_details = False
self.mhstats = {}
self.made_proposals = 0
self.accepted_proposals = 0
self.hashval = rrandom.random.randbelow()
return
def mhstats_on(self):
self.mhstats_details = True
def mhstats_off(self):
self.mhstats_details = False
def mhstats_aggregated(self):
d = {}
d['made-proposals'] = self.made_proposals
d['accepted-proposals'] = self.accepted_proposals
return d
def mhstats_detailed(self):
return self.mhstats
def get_log_score(self, id):
if id == -1:
return self.p
else:
return self.observes[id].p
def weight(self):
return self.db.weight() + self.weighted_db.weight()
def random_choices(self):
return self.db.__len__() + self.weighted_db.__len__()
def assume(self, name, expr, id=-1):
evalnode = ReducedEvalNode(self, self.env, expr)
self.env.add_assume(name, evalnode)
evalnode.add_assume(name, self.env)
if id != -1:
self.assumes[id] = evalnode
assert id == len(self.directives)
self.directives.append('assume')
val = evalnode.evaluate()
return val
def predict(self, expr, id):
evalnode = ReducedEvalNode(self, self.env, expr)
assert id == len(self.directives)
self.directives.append('predict')
self.predicts[id] = evalnode
evalnode.predict = True
val = evalnode.evaluate()
return val
def observe(self, expr, obs_val, id):
evalnode = ReducedEvalNode(self, self.env, expr)
assert id == len(self.directives)
self.directives.append('observe')
self.observes[id] = evalnode
evalnode.observed = True
evalnode.observe_val = obs_val
val = evalnode.evaluate()
return val
def forget(self, id):
if id in self.observes:
d = self.observes
elif id in self.predicts:
d = self.predicts
else:
raise RException("Can only forget predicts and observes")
evalnode = d[id]
evalnode.unevaluate()
#del d[id]
return
def report_value(self, id):
node = self.get_directive_node(id)
if not node.active:
raise RException("Error. Perhaps this directive was forgotten?")
val = node.val
return val
def get_directive_node(self, id):
if self.directives[id] == 'assume':
node = self.assumes[id]
elif self.directives[id] == 'observe':
node = self.observes[id]
else:
assert self.directives[id] == 'predict'
node = self.predicts[id]
return node
def add_accepted_proposal(self, hashval):
if self.mhstats_details:
if hashval in self.mhstats:
self.mhstats[hashval]['accepted-proposals'] += 1
else:
self.mhstats[hashval] = {}
self.mhstats[hashval]['accepted-proposals'] = 1
self.mhstats[hashval]['made-proposals'] = 0
self.accepted_proposals += 1
def add_made_proposal(self, hashval):
if self.mhstats_details:
if hashval in self.mhstats:
self.mhstats[hashval]['made-proposals'] += 1
else:
self.mhstats[hashval] = {}
self.mhstats[hashval]['made-proposals'] = 1
self.mhstats[hashval]['accepted-proposals'] = 0
self.made_proposals += 1
def reflip(self, hashval):
if self.debug:
print self
if hashval in self.choices:
self.application_reflip = True
self.reflip_node = self.choices[hashval]
if not self.reflip_node.random_xrp_apply:
raise RException(
"Reflipping something which isn't a random xrp application"
)
if self.reflip_node.val is None:
raise RException(
"Reflipping something which previously had value None")
else:
self.application_reflip = False # internal reflip
(self.reflip_xrp, nodes) = self.xrps[hashval]
self.nodes = list_nodes(nodes)
self.eval_p = 0
self.uneval_p = 0
old_p = self.p
self.old_to_new_q = -math.log(self.weight())
if self.application_reflip:
self.old_val = self.reflip_node.val
self.new_val = self.reflip_node.reflip()
else:
# TODO: this is copied from traces. is it correct?
args_list = []
self.old_vals = []
for node in self.nodes:
args_list.append(node.args)
self.old_vals.append(node.val)
self.old_to_new_q += math.log(self.reflip_xrp.state_weight())
old_p += self.reflip_xrp.theta_prob()
self.new_vals, q_forwards, q_back = self.reflip_xrp.theta_mh_prop(
args_list, self.old_vals)
self.old_to_new_q += q_forwards
self.new_to_old_q += q_back
for i in range(len(self.nodes)):
node = self.nodes[i]
val = self.new_vals[i]
node.set_val(val)
node.propagate_up(False)
new_p = self.p
self.new_to_old_q = -math.log(self.weight())
self.old_to_new_q += self.eval_p
self.new_to_old_q += self.uneval_p
if not self.application_reflip:
new_p += self.reflip_xrp.theta_prob()
self.new_to_old_q += math.log(self.reflip_xrp.state_weight())
if self.debug:
if self.application_reflip:
print "\nCHANGING VAL OF ", self.reflip_node, "\n FROM : ", self.old_val, "\n TO : ", self.new_val, "\n"
if (self.old_val.__eq__(self.new_val)).bool:
print "SAME VAL"
else:
print "TRANSITIONING STATE OF ", self.reflip_xrp
print "new db", self
print "\nq(old -> new) : ", math.exp(self.old_to_new_q)
print "q(new -> old) : ", math.exp(self.new_to_old_q)
print "p(old) : ", math.exp(old_p)
print "p(new) : ", math.exp(new_p)
print 'transition prob : ', math.exp(new_p + self.new_to_old_q -
old_p -
self.old_to_new_q), "\n"
print "\n-----------------------------------------\n"
return old_p, self.old_to_new_q, new_p, self.new_to_old_q
def restore(self):
if self.application_reflip:
self.reflip_node.restore(self.old_val)
else:
for i in range(len(self.nodes)):
node = self.nodes[i]
node.set_val(self.old_vals[i])
node.propagate_up(True)
self.reflip_xrp.theta_mh_restore()
if self.debug:
print 'restore'
def keep(self):
if self.application_reflip:
self.reflip_node.restore(
self.new_val
) # NOTE: Is necessary for correctness, as we must forget old branch
else:
for i in range(len(self.nodes)):
node = self.nodes[i]
node.restore(self.new_vals[i])
self.reflip_xrp.theta_mh_keep()
def add_for_transition(self, xrp, evalnode):
hashval = xrp.__hash__()
if hashval not in self.xrps:
self.xrps[hashval] = (xrp, {})
(xrp, evalnodes) = self.xrps[hashval]
evalnodes[evalnode] = True
weight = xrp.state_weight()
self.delete_from_db(xrp.__hash__())
self.add_to_db(xrp.__hash__(), weight)
# Add an XRP application node to the db
def add_xrp(self, xrp, args, evalnode):
weight = xrp.weight(args)
evalnode.setargs(args)
try:
self.new_to_old_q += math.log(weight)
except:
pass # This is only necessary if we're reflipping
if self.weighted_db.__contains__(
evalnode.hashval) or self.db.__contains__(
evalnode.hashval) or (evalnode.hashval in self.choices):
raise RException("DB already had this evalnode")
self.choices[evalnode.hashval] = evalnode
self.add_to_db(evalnode.hashval, weight)
def add_to_db(self, hashval, weight):
if weight == 0:
return
elif weight == 1:
self.db.__setitem__(hashval, True)
else:
self.weighted_db.__setitem__(hashval, True, weight)
def remove_for_transition(self, xrp, evalnode):
hashval = xrp.__hash__()
(xrp, evalnodes) = self.xrps[hashval]
del evalnodes[evalnode]
if len(evalnodes) == 0:
del self.xrps[hashval]
self.delete_from_db(xrp.__hash__())
def remove_xrp(self, evalnode):
xrp = evalnode.xrp
try:
self.old_to_new_q += math.log(xrp.weight(evalnode.args))
except:
pass # This fails when restoring/keeping, for example
if evalnode.hashval not in self.choices:
raise RException("Choices did not already have this evalnode")
else:
del self.choices[evalnode.hashval]
self.delete_from_db(evalnode.hashval)
def delete_from_db(self, hashval):
if self.db.__contains__(hashval):
self.db.__delitem__(hashval)
elif self.weighted_db.__contains__(hashval):
self.weighted_db.__delitem__(hashval)
def randomKey(self):
if rrandom.random.random() * self.weight() > self.db.weight():
return self.weighted_db.randomKey()
else:
return self.db.randomKey()
def infer(self):
try:
hashval = self.randomKey()
except:
raise RException("Program has no randomness!")
old_p, old_to_new_q, new_p, new_to_old_q = self.reflip(hashval)
p = rrandom.random.random()
if new_p + new_to_old_q - old_p - old_to_new_q < math.log(p):
self.restore()
else:
self.keep()
self.add_accepted_proposal(hashval)
self.add_made_proposal(hashval)
def reset(self):
self.__init__()
def __str__(self):
string = "EvalNodeTree:"
for evalnode in self.assumes.values():
string += evalnode.str_helper()
for evalnode in self.observes.values():
string += evalnode.str_helper()
for evalnode in self.predicts.values():
string += evalnode.str_helper()
return string
class ReducedTraces(Engine):
def __init__(self):
self.engine_type = 'reduced traces'
self.assumes = {} # id -> evalnode
self.observes = {} # id -> evalnode
self.predicts = {} # id -> evalnode
self.directives = []
self.db = RandomChoiceDict()
self.weighted_db = WeightedRandomChoiceDict()
self.choices = {} # hash -> evalnode
self.xrps = {} # hash -> (xrp, set of application nodes)
self.env = EnvironmentNode()
self.p = 0
self.uneval_p = 0
self.eval_p = 0
self.new_to_old_q = 0
self.old_to_new_q = 0
self.debug = False
# necessary because of the new XRP interface requiring some state kept while doing inference
self.application_reflip = False
self.reflip_node = ReducedEvalNode(self, self.env, VarExpression(''))
self.nodes = []
self.old_vals = [Value()]
self.new_vals = [Value()]
self.old_val = Value()
self.new_val = Value()
self.reflip_xrp = XRP()
self.mhstats_details = False
self.mhstats = {}
self.made_proposals = 0
self.accepted_proposals = 0
self.hashval = rrandom.random.randbelow()
return
def mhstats_on(self):
self.mhstats_details = True
def mhstats_off(self):
self.mhstats_details = False
def mhstats_aggregated(self):
d = {}
d['made-proposals'] = self.made_proposals
d['accepted-proposals'] = self.accepted_proposals
return d
def mhstats_detailed(self):
return self.mhstats
def get_log_score(self, id):
if id == -1:
return self.p
else:
return self.observes[id].p
def weight(self):
return self.db.weight() + self.weighted_db.weight()
def random_choices(self):
return self.db.__len__() + self.weighted_db.__len__()
def assume(self, name, expr, id = -1):
evalnode = ReducedEvalNode(self, self.env, expr)
self.env.add_assume(name, evalnode)
evalnode.add_assume(name, self.env)
if id != -1:
self.assumes[id] = evalnode
assert id == len(self.directives)
self.directives.append('assume')
val = evalnode.evaluate()
return val
def predict(self, expr, id):
evalnode = ReducedEvalNode(self, self.env, expr)
assert id == len(self.directives)
self.directives.append('predict')
self.predicts[id] = evalnode
evalnode.predict = True
val = evalnode.evaluate()
return val
def observe(self, expr, obs_val, id):
evalnode = ReducedEvalNode(self, self.env, expr)
assert id == len(self.directives)
self.directives.append('observe')
self.observes[id] = evalnode
evalnode.observed = True
evalnode.observe_val = obs_val
val = evalnode.evaluate()
return val
def forget(self, id):
if id in self.observes:
d = self.observes
elif id in self.predicts:
d = self.predicts
else:
raise RException("Can only forget predicts and observes")
evalnode = d[id]
evalnode.unevaluate()
#del d[id]
return
def report_value(self, id):
node = self.get_directive_node(id)
if not node.active:
raise RException("Error. Perhaps this directive was forgotten?")
val = node.val
return val
def get_directive_node(self, id):
if self.directives[id] == 'assume':
node = self.assumes[id]
elif self.directives[id] == 'observe':
node = self.observes[id]
else:
assert self.directives[id] == 'predict'
node = self.predicts[id]
return node
def add_accepted_proposal(self, hashval):
if self.mhstats_details:
if hashval in self.mhstats:
self.mhstats[hashval]['accepted-proposals'] += 1
else:
self.mhstats[hashval] = {}
self.mhstats[hashval]['accepted-proposals'] = 1
self.mhstats[hashval]['made-proposals'] = 0
self.accepted_proposals += 1
def add_made_proposal(self, hashval):
if self.mhstats_details:
if hashval in self.mhstats:
self.mhstats[hashval]['made-proposals'] += 1
else:
self.mhstats[hashval] = {}
self.mhstats[hashval]['made-proposals'] = 1
self.mhstats[hashval]['accepted-proposals'] = 0
self.made_proposals += 1
def reflip(self, hashval):
if self.debug:
print self
if hashval in self.choices:
self.application_reflip = True
self.reflip_node = self.choices[hashval]
if not self.reflip_node.random_xrp_apply:
raise RException("Reflipping something which isn't a random xrp application")
if self.reflip_node.val is None:
raise RException("Reflipping something which previously had value None")
else:
self.application_reflip = False # internal reflip
(self.reflip_xrp, nodes) = self.xrps[hashval]
self.nodes = list_nodes(nodes)
self.eval_p = 0
self.uneval_p = 0
old_p = self.p
self.old_to_new_q = - math.log(self.weight())
if self.application_reflip:
self.old_val = self.reflip_node.val
self.new_val = self.reflip_node.reflip()
else:
# TODO: this is copied from traces. is it correct?
args_list = []
self.old_vals = []
for node in self.nodes:
args_list.append(node.args)
self.old_vals.append(node.val)
self.old_to_new_q += math.log(self.reflip_xrp.state_weight())
old_p += self.reflip_xrp.theta_prob()
self.new_vals, q_forwards, q_back = self.reflip_xrp.theta_mh_prop(args_list, self.old_vals)
self.old_to_new_q += q_forwards
self.new_to_old_q += q_back
for i in range(len(self.nodes)):
node = self.nodes[i]
val = self.new_vals[i]
node.set_val(val)
node.propagate_up(False)
new_p = self.p
self.new_to_old_q = - math.log(self.weight())
self.old_to_new_q += self.eval_p
self.new_to_old_q += self.uneval_p
if not self.application_reflip:
new_p += self.reflip_xrp.theta_prob()
self.new_to_old_q += math.log(self.reflip_xrp.state_weight())
if self.debug:
if self.application_reflip:
print "\nCHANGING VAL OF ", self.reflip_node, "\n FROM : ", self.old_val, "\n TO : ", self.new_val, "\n"
if (self.old_val.__eq__(self.new_val)).bool:
print "SAME VAL"
else:
print "TRANSITIONING STATE OF ", self.reflip_xrp
print "new db", self
print "\nq(old -> new) : ", math.exp(self.old_to_new_q)
print "q(new -> old) : ", math.exp(self.new_to_old_q )
print "p(old) : ", math.exp(old_p)
print "p(new) : ", math.exp(new_p)
print 'transition prob : ', math.exp(new_p + self.new_to_old_q - old_p - self.old_to_new_q) , "\n"
print "\n-----------------------------------------\n"
return old_p, self.old_to_new_q, new_p, self.new_to_old_q
def restore(self):
if self.application_reflip:
self.reflip_node.restore(self.old_val)
else:
for i in range(len(self.nodes)):
node = self.nodes[i]
node.set_val(self.old_vals[i])
node.propagate_up(True)
self.reflip_xrp.theta_mh_restore()
if self.debug:
print 'restore'
def keep(self):
if self.application_reflip:
self.reflip_node.restore(self.new_val) # NOTE: Is necessary for correctness, as we must forget old branch
else:
for i in range(len(self.nodes)):
node = self.nodes[i]
node.restore(self.new_vals[i])
self.reflip_xrp.theta_mh_keep()
def add_for_transition(self, xrp, evalnode):
hashval = xrp.__hash__()
if hashval not in self.xrps:
self.xrps[hashval] = (xrp, {})
(xrp, evalnodes) = self.xrps[hashval]
evalnodes[evalnode] = True
weight = xrp.state_weight()
self.delete_from_db(xrp.__hash__())
self.add_to_db(xrp.__hash__(), weight)
# Add an XRP application node to the db
def add_xrp(self, xrp, args, evalnode):
weight = xrp.weight(args)
evalnode.setargs(args)
try:
self.new_to_old_q += math.log(weight)
except:
pass # This is only necessary if we're reflipping
if self.weighted_db.__contains__(evalnode.hashval) or self.db.__contains__(evalnode.hashval) or (evalnode.hashval in self.choices):
raise RException("DB already had this evalnode")
self.choices[evalnode.hashval] = evalnode
self.add_to_db(evalnode.hashval, weight)
def add_to_db(self, hashval, weight):
if weight == 0:
return
elif weight == 1:
self.db.__setitem__(hashval, True)
else:
self.weighted_db.__setitem__(hashval, True, weight)
def remove_for_transition(self, xrp, evalnode):
hashval = xrp.__hash__()
(xrp, evalnodes) = self.xrps[hashval]
del evalnodes[evalnode]
if len(evalnodes) == 0:
del self.xrps[hashval]
self.delete_from_db(xrp.__hash__())
def remove_xrp(self, evalnode):
xrp = evalnode.xrp
try:
self.old_to_new_q += math.log(xrp.weight(evalnode.args))
except:
pass # This fails when restoring/keeping, for example
if evalnode.hashval not in self.choices:
raise RException("Choices did not already have this evalnode")
else:
del self.choices[evalnode.hashval]
self.delete_from_db(evalnode.hashval)
def delete_from_db(self, hashval):
if self.db.__contains__(hashval):
self.db.__delitem__(hashval)
elif self.weighted_db.__contains__(hashval):
self.weighted_db.__delitem__(hashval)
def randomKey(self):
if rrandom.random.random() * self.weight() > self.db.weight():
return self.weighted_db.randomKey()
else:
return self.db.randomKey()
def infer(self):
try:
hashval = self.randomKey()
except:
raise RException("Program has no randomness!")
old_p, old_to_new_q, new_p, new_to_old_q = self.reflip(hashval)
p = rrandom.random.random()
if new_p + new_to_old_q - old_p - old_to_new_q < math.log(p):
self.restore()
else:
self.keep()
self.add_accepted_proposal(hashval)
self.add_made_proposal(hashval)
def reset(self):
self.__init__()
def __str__(self):
string = "EvalNodeTree:"
for evalnode in self.assumes.values():
string += evalnode.str_helper()
for evalnode in self.observes.values():
string += evalnode.str_helper()
for evalnode in self.predicts.values():
string += evalnode.str_helper()
return string
class RandomDB(Engine):
def __init__(self):
#self.db = {}
self.engine_type = 'randomdb'
self.db = RandomChoiceDict()
self.db_noise = {}
self.log = []
# ALWAYS WORKING WITH LOG PROBABILITIES
self.uneval_p = 0
self.eval_p = 0
self.p = 0
self.env = Environment()
self.assumes = {}
self.observes = {}
self.predicts = {}
self.vars = {}
def reset(self):
self.__init__()
def assume(self, varname, expr, id):
self.assumes[id] = (varname, expr)
self.vars[varname] = expr
value = self.evaluate(expr, self.env, reflip=True, stack=[id])
self.env.set(varname, value)
return value
def observe(self, expr, obs_val, id):
if expr.hashval in self.observes:
raise RException('Already observed %s' % str(expr))
self.observes[id] = (expr, obs_val)
# bit of a hack, here, to make it recognize same things as with noisy_expr
self.evaluate(expr,
self.env,
reflip=False,
stack=[id],
xrp_force_val=obs_val)
return expr.hashval
def predict(self, expr, id):
self.predicts[id] = expr
return self.evaluate(expr, self.env, True, [id])
def forget(self, id):
self.remove(['obs', id])
assert id in self.observes
del self.observes[id]
return
def insert(self,
stack,
xrp,
value,
args,
is_obs_noise=False,
memorize=True):
stack = tuple(stack)
if self.has(stack):
self.remove(stack)
prob = xrp.prob(value, args)
self.p += prob
xrp.incorporate(value, args)
if is_obs_noise:
self.db_noise[stack] = (xrp, value, args, True)
else:
self.db[stack] = (xrp, value, args, False)
if not is_obs_noise:
self.eval_p += prob # hmmm..
if memorize:
self.log.append(('insert', stack, xrp, value, args, is_obs_noise))
def remove(self, stack, memorize=True):
stack = tuple(stack)
assert self.has(stack)
(xrp, value, args, is_obs_noise) = self.get(stack)
xrp.remove(value, args)
prob = xrp.prob(value, args)
self.p -= prob
if is_obs_noise:
del self.db_noise[stack]
else:
del self.db[stack]
self.uneval_p += prob # previously unindented...
if memorize:
self.log.append(('remove', stack, xrp, value, args, is_obs_noise))
def has(self, stack):
stack = tuple(stack)
return ((stack in self.db) or (stack in self.db_noise))
def get(self, stack):
stack = tuple(stack)
if stack in self.db:
return self.db[stack]
elif stack in self.db_noise:
return self.db_noise[stack]
else:
raise RException('Failed to get stack %s' % str(stack))
def infer(self):
try:
stack = self.random_stack()
except:
return
self.reflip(stack)
def random_stack(self):
key = self.db.randomKey()
return key
def evaluate_recurse(self, subexpr, env, reflip, stack, addition):
newstack = [s for s in stack].extend(addition)
val = self.evaluate(subexpr, env, reflip, newstack)
return val
def binary_op_evaluate(self, expr, env, reflip, stack):
val1 = self.evaluate_recurse(expr.children[0], env, reflip, stack,
['operand0'])
val2 = self.evaluate_recurse(expr.children[1], env, reflip, stack,
['operand1'])
return (val1, val2)
def children_evaluate(self, expr, env, reflip, stack):
return [
self.evaluate_recurse(expr.children[i], env, reflip, stack,
['child' + str(i)])
for i in range(len(expr.children))
]
# Draws a sample value (without re-sampling other values) given its parents, and sets it
def evaluate(self,
expr,
env=None,
reflip=False,
stack=[],
xrp_force_val=None):
if env is None:
env = self.env
if xrp_force_val is not None:
assert expr.type == 'apply'
if expr.type == 'value':
val = expr.val
elif expr.type == 'variable':
var = expr.name
(val, lookup_env) = env.lookup(var)
elif expr.type == 'if':
cond = self.evaluate_recurse(expr.cond, env, reflip, stack,
['cond'])
if cond.bool:
self.unevaluate(stack + ['false'])
val = self.evaluate_recurse(expr.true, env, reflip, stack,
['true'])
else:
self.unevaluate(stack + ['true'])
val = self.evaluate_recurse(expr.false, env, reflip, stack,
['false'])
#elif expr.type == 'switch':
# index = self.evaluate_recurse(expr.index, env, reflip, stack , ['index'])
# assert 0 <= index.num < expr.n
# for i in range(expr.n):
# if i != index.num:
# self.unevaluate(stack + ['child' + str(i)])
# val = self.evaluate_recurse(expr.children[index.num], env, reflip, stack, ['child' + str(index.num)])
elif expr.type == 'let':
# NOTE: this really is a let*
n = len(expr.vars)
assert len(expr.expressions) == n
values = []
new_env = env
for i in range(n): # Bind variables
new_env = new_env.spawn_child()
val = self.evaluate_recurse(expr.expressions[i], new_env,
reflip, stack, ['let' + str(i)])
values.append(val)
new_env.set(expr.vars[i], values[i])
if val.type == 'procedure':
val.env = new_env
new_body = expr.body.replace(new_env)
val = self.evaluate_recurse(new_body, new_env, reflip, stack,
['body'])
elif expr.type == 'apply':
n = len(expr.children)
args = [
self.evaluate_recurse(expr.children[i], env, reflip, stack,
['arg' + str(i)]) for i in range(n)
]
op = self.evaluate_recurse(expr.op, env, reflip, stack,
['operator'])
addition = ','.join([x.str_hash for x in args])
if op.type == 'procedure':
self.unevaluate(stack + ['apply'], addition)
if n != len(op.vars):
raise RException(
'Procedure should have %d arguments. \nVars were \n%s\n, but children were \n%s.'
% (n, op.vars, expr.children))
new_env = op.env.spawn_child()
for i in range(n):
new_env.set(op.vars[i], args[i])
val = self.evaluate_recurse(op.body, new_env, reflip, stack,
['apply', addition])
elif op.type == 'xrp':
self.unevaluate(stack + ['apply'], addition)
if xrp_force_val is not None:
assert not reflip
if self.has(stack):
self.remove(stack)
self.insert(stack, op.xrp, xrp_force_val, args, True)
val = xrp_force_val
else:
substack = stack + ['apply', addition]
if not self.has(substack):
if op.xrp.is_mem_proc():
val = op.xrp.apply_mem(args, self, stack)
else:
val = op.xrp.apply(args)
self.insert(substack, op.xrp, val, args)
else:
if reflip:
self.remove(substack)
if op.xrp.is_mem_proc():
val = op.xrp.apply_mem(args, self, stack)
else:
val = op.xrp.apply(args)
self.insert(substack, op.xrp, val, args)
else:
(xrp, val, dbargs,
is_obs_noise) = self.get(substack)
assert not is_obs_noise
else:
raise RException(
'Must apply either a procedure or xrp. Instead got expression %s'
% str(op))
elif expr.type == 'function':
n = len(expr.vars)
new_env = env.spawn_child()
bound = {}
for i in range(n): # Bind variables
bound[expr.vars[i]] = True
procedure_body = expr.body.replace(new_env, bound)
val = Procedure(expr.vars, procedure_body, env)
elif expr.type == '=':
(val1, val2) = self.binary_op_evaluate(expr, env, reflip, stack)
val = val1.__eq__(val2)
elif expr.type == '<':
(val1, val2) = self.binary_op_evaluate(expr, env, reflip, stack)
val = val1.__lt__(val2)
elif expr.type == '>':
(val1, val2) = self.binary_op_evaluate(expr, env, reflip, stack)
val = val1.__gt__(val2)
elif expr.type == '<=':
(val1, val2) = self.binary_op_evaluate(expr, env, reflip, stack)
val = val1.__le__(val2)
elif expr.type == '>=':
(val1, val2) = self.binary_op_evaluate(expr, env, reflip, stack)
val = val1.__ge__(val2)
elif expr.type == '&':
vals = self.children_evaluate(expr, env, reflip, stack)
andval = BoolValue(True)
for x in vals:
andval = andval.__and__(x.bool)
val = andval
elif expr.type == '^':
vals = self.children_evaluate(expr, env, reflip, stack)
xorval = BoolValue(True)
for x in vals:
xorval = xorval.__xor__(x.bool)
val = xorval
elif expr.type == '|':
vals = self.children_evaluate(expr, env, reflip, stack)
orval = BoolValue(False)
for x in vals:
orval = orval.__or__(x.bool)
val = orval
elif expr.type == '~':
negval = self.evaluate_recurse(expr.children[0], env, reflip,
stack, ['neg'])
val = negval.__inv__()
elif expr.type == '+':
vals = self.children_evaluate(expr, env, reflip, stack)
sum_val = NatValue(0)
for x in vals:
sum_val = sum_val.__add__(x)
val = sum_val
elif expr.type == '-':
val1 = self.evaluate_recurse(expr.children[0], env, reflip, stack,
['sub0'])
val2 = self.evaluate_recurse(expr.children[1], env, reflip, stack,
['sub1'])
val = val1.__sub__(val2)
elif expr.type == '*':
vals = self.children_evaluate(expr, env, reflip, stack)
prod_val = NatValue(1)
for x in vals:
prod_val = prod_val.__mul__(x)
val = prod_val
elif expr.type == '/':
val1 = self.evaluate_recurse(expr.children[0], env, reflip, stack,
['div0'])
val2 = self.evaluate_recurse(expr.children[1], env, reflip, stack,
['div1'])
val = val1.__div__(val2)
else:
raise RException('Invalid expression type %s' % expr.type)
return val
def unevaluate(self, uneval_stack, args=None):
if args is not None:
args = tuple(args)
to_unevaluate = []
for tuple_stack in self.db:
to_unevaluate.append(tuple_stack)
for tuple_stack in self.db_noise:
to_unevaluate.append(tuple_stack)
to_delete = []
for tuple_stack in to_unevaluate:
stack = list(tuple_stack)
if len(stack) >= len(uneval_stack) and stack[:len(
uneval_stack)] == uneval_stack:
if args is None:
to_delete.append(tuple_stack)
else:
assert len(stack) > len(uneval_stack)
if stack[len(uneval_stack)] != args:
to_delete.append(tuple_stack)
for tuple_stack in to_delete:
self.remove(tuple_stack)
def save(self):
self.log = []
self.uneval_p = 0
self.eval_p = 0
def restore(self):
self.log.reverse()
for (type, stack, xrp, value, args, is_obs_noise) in self.log:
if type == 'insert':
self.remove(stack, False)
else:
assert type == 'remove'
self.insert(stack, xrp, value, args, is_obs_noise, False)
def reflip(self, stack):
(xrp, val, args, is_obs_noise) = self.get(stack)
#debug = True
debug = False
old_p = self.p
old_to_new_q = -math.log(len(self.db))
if debug:
print "old_db", self
self.save()
self.remove(stack)
if xrp.is_mem_proc():
new_val = xrp.apply(args, list(stack))
else:
new_val = xrp.apply(args)
self.insert(stack, xrp, new_val, args)
if debug:
print "\nCHANGING ", stack, "\n TO : ", new_val, "\n"
if val == new_val:
return
self.rerun(False)
new_p = self.p
new_to_old_q = self.uneval_p - math.log(len(self.db))
old_to_new_q += self.eval_p
if debug:
print "new db", self, \
"\nq(old -> new) : ", old_to_new_q, \
"q(new -> old) : ", new_to_old_q, \
"p(old) : ", old_p, \
"p(new) : ", new_p, \
'log transition prob : ', new_p + new_to_old_q - old_p - old_to_new_q , "\n"
if old_p * old_to_new_q > 0:
p = rrandom.random.random()
if new_p + new_to_old_q - old_p - old_to_new_q < math.log(p):
self.restore()
if debug:
print 'restore'
self.rerun(False)
if debug:
print "new db", self
print "\n-----------------------------------------\n"
def __str__(self):
string = 'DB with state:'
string += '\n Regular Flips:'
for s in self.db:
string += '\n %s <- %s' % (self.db[s][1].val, s)
string += '\n Observe Flips:'
for s in self.db_noise:
string += '\n %s <- %s' % (self.db_noise[s][1].val, s)
return string
def __contains__(self, stack):
return self.has(self, stack)
def __getitem__(self, stack):
return self.get(self, stack)
class Traces(Engine):
def __init__(self):
self.engine_type = "traces"
self.assumes = {} # id -> evalnode
self.observes = {} # id -> evalnode
self.predicts = {} # id -> evalnode
self.directives = []
self.db = RandomChoiceDict()
self.weighted_db = WeightedRandomChoiceDict()
self.choices = {} # hash -> evalnode
self.xrps = {} # hash -> (xrp, set of application nodes)
self.env = EnvironmentNode()
self.p = 0
self.old_to_new_q = 0
self.new_to_old_q = 0
self.eval_xrps = [] # (xrp, args, val)
self.uneval_xrps = [] # (xrp, args, val)
self.debug = False
# Stuff for restoring
self.application_reflip = False
self.reflip_node = EvalNode(self, self.env, VarExpression(""))
self.nodes = []
self.old_vals = []
self.new_vals = []
self.old_val = Value()
self.new_val = Value()
self.reflip_xrp = XRP()
self.made_proposals = 0
self.accepted_proposals = 0
self.mhstats_details = False
self.mhstats = {}
return
def mhstats_on(self):
self.mhstats_details = True
def mhstats_off(self):
self.mhstats_details = False
def mhstats_aggregated(self):
d = {}
d["made-proposals"] = self.made_proposals
d["accepted-proposals"] = self.accepted_proposals
return d
def mhstats_detailed(self):
return self.mhstats
def get_log_score(self, id):
if id == -1:
return self.p
else:
return self.observes[id].p
def weight(self):
return self.db.weight() + self.weighted_db.weight()
def random_choices(self):
return self.db.__len__() + self.weighted_db.__len__()
def assume(self, name, expr, id):
evalnode = EvalNode(self, self.env, expr)
if id != -1:
self.assumes[id] = evalnode
if id != len(self.directives):
raise RException("Id %d does not agree with directives length of %d" % (id, len(self.directives)))
self.directives.append("assume")
val = evalnode.evaluate()
self.env.add_assume(name, evalnode)
evalnode.add_assume(name, self.env)
self.env.set(name, val)
return val
def predict(self, expr, id):
evalnode = EvalNode(self, self.env, expr)
if id != len(self.directives):
raise RException("Id %d does not agree with directives length of %d" % (id, len(self.directives)))
self.directives.append("predict")
self.predicts[id] = evalnode
val = evalnode.evaluate(False)
return val
def observe(self, expr, obs_val, id):
evalnode = EvalNode(self, self.env, expr)
if id != len(self.directives):
raise RException("Id %d does not agree with directives length of %d" % (id, len(self.directives)))
self.directives.append("observe")
self.observes[id] = evalnode
evalnode.observed = True
evalnode.observe_val = obs_val
val = evalnode.evaluate()
return val
def forget(self, id):
if id in self.observes:
d = self.observes
elif id in self.predicts:
d = self.predicts
else:
raise RException("Can only forget predicts and observes")
evalnode = d[id]
evalnode.unevaluate()
# del d[id]
return
def report_value(self, id):
node = self.get_directive_node(id)
if not node.active:
raise RException("Error. Perhaps this directive was forgotten?")
val = node.val
return val
def get_directive_node(self, id):
if self.directives[id] == "assume":
node = self.assumes[id]
elif self.directives[id] == "observe":
node = self.observes[id]
elif self.directives[id] == "predict":
node = self.predicts[id]
else:
raise RException("Invalid directive")
return node
def add_accepted_proposal(self, hashval):
if self.mhstats_details:
if hashval in self.mhstats:
self.mhstats[hashval]["accepted-proposals"] += 1
else:
self.mhstats[hashval] = {}
self.mhstats[hashval]["accepted-proposals"] = 1
self.mhstats[hashval]["made-proposals"] = 0
self.accepted_proposals += 1
def add_made_proposal(self, hashval):
if self.mhstats_details:
if hashval in self.mhstats:
self.mhstats[hashval]["made-proposals"] += 1
else:
self.mhstats[hashval] = {}
self.mhstats[hashval]["made-proposals"] = 1
self.mhstats[hashval]["accepted-proposals"] = 0
self.made_proposals += 1
def reflip(self, hashval):
if self.debug:
print self
if hashval in self.choices:
self.application_reflip = True
self.reflip_node = self.choices[hashval]
if not self.reflip_node.random_xrp_apply:
raise RException("Reflipping something which isn't a random xrp application")
if self.reflip_node.val is None:
raise RException("Reflipping something which previously had value None")
else:
self.application_reflip = False # internal reflip
(self.reflip_xrp, nodes) = self.xrps[hashval]
self.nodes = list_nodes(nodes)
self.old_to_new_q = 0
self.new_to_old_q = 0
old_p = self.p
self.old_to_new_q = -math.log(self.weight())
if self.application_reflip:
self.old_val = self.reflip_node.val
self.new_val = self.reflip_node.reflip()
else:
args_list = []
self.old_vals = []
for node in self.nodes:
if not node.xrp_apply:
raise RException("non-XRP application node being used in transition")
args_list.append(node.args)
self.old_vals.append(node.val)
self.old_to_new_q += math.log(self.reflip_xrp.state_weight())
old_p += self.reflip_xrp.theta_prob()
self.new_vals, q_forwards, q_back = self.reflip_xrp.theta_mh_prop(args_list, self.old_vals)
self.old_to_new_q += q_forwards
self.new_to_old_q += q_back
for i in range(len(self.nodes)):
node = self.nodes[i]
val = self.new_vals[i]
node.set_val(val)
node.propagate_up(False, True)
new_p = self.p
self.new_to_old_q -= math.log(self.weight())
if not self.application_reflip:
new_p += self.reflip_xrp.theta_prob()
self.new_to_old_q += math.log(self.reflip_xrp.state_weight())
if self.debug:
if self.application_reflip:
print "\nCHANGING VAL OF ", self.reflip_node, "\n FROM : ", self.old_val, "\n TO : ", self.new_val, "\n"
if (self.old_val.__eq__(self.new_val)).bool:
print "SAME VAL"
else:
print "TRANSITIONING STATE OF ", self.reflip_xrp
print "new db", self
print "\nq(old -> new) : ", math.exp(self.old_to_new_q)
print "q(new -> old) : ", math.exp(self.new_to_old_q)
print "p(old) : ", math.exp(old_p)
print "p(new) : ", math.exp(new_p)
print "transition prob : ", math.exp(new_p + self.new_to_old_q - old_p - self.old_to_new_q), "\n"
return old_p, self.old_to_new_q, new_p, self.new_to_old_q
def restore(self):
if self.application_reflip:
self.reflip_node.restore(self.old_val)
else:
for i in range(len(self.nodes)):
node = self.nodes[i]
node.set_val(self.old_vals[i])
node.propagate_up(True, True)
self.reflip_xrp.theta_mh_restore()
if self.debug:
print "restore"
print "\n-----------------------------------------\n"
def keep(self):
if self.application_reflip:
self.reflip_node.restore(self.new_val) # NOTE: Is necessary for correctness, as we must forget old branch
else:
for i in range(len(self.nodes)):
node = self.nodes[i]
node.restore(self.new_vals[i])
self.reflip_xrp.theta_mh_keep()
if self.debug:
print "keep"
print "\n-----------------------------------------\n"
def add_for_transition(self, xrp, evalnode):
hashval = xrp.__hash__()
if hashval not in self.xrps:
self.xrps[hashval] = (xrp, {})
(xrp, evalnodes) = self.xrps[hashval]
evalnodes[evalnode] = True
weight = xrp.state_weight()
self.delete_from_db(hashval)
self.add_to_db(hashval, weight)
# Add an XRP application node to the db
def add_xrp(self, xrp, args, val, evalnode, forcing=False):
self.eval_xrps.append((xrp, args, val))
evalnode.forcing = forcing
if not xrp.resample:
evalnode.random_xrp_apply = True
prob = xrp.prob(val, args)
evalnode.setargs(xrp, args, prob)
xrp.incorporate(val, args)
xrp.make_link(evalnode, args)
if not forcing:
self.old_to_new_q += prob
self.p += prob
self.add_for_transition(xrp, evalnode)
if xrp.resample:
return
weight = xrp.weight(args)
try:
self.new_to_old_q += math.log(weight)
except:
pass # This is only necessary if we're reflipping
if (
self.weighted_db.__contains__(evalnode.hashval)
or self.db.__contains__(evalnode.hashval)
or (evalnode.hashval in self.choices)
):
raise RException("DB already had this evalnode")
self.choices[evalnode.hashval] = evalnode
self.add_to_db(evalnode.hashval, weight)
def add_to_db(self, hashval, weight):
if weight == 0:
return
elif weight == 1:
self.db.__setitem__(hashval, True)
else:
self.weighted_db.__setitem__(hashval, True, weight)
def remove_for_transition(self, xrp, evalnode):
hashval = xrp.__hash__()
(xrp, evalnodes) = self.xrps[hashval]
del evalnodes[evalnode]
if len(evalnodes) == 0:
del self.xrps[hashval]
self.delete_from_db(hashval)
def remove_xrp(self, xrp, args, val, evalnode):
self.uneval_xrps.append((xrp, args, val))
xrp.break_link(evalnode, args)
xrp.remove(val, args)
prob = xrp.prob(val, args)
if not evalnode.forcing:
self.new_to_old_q += prob
self.p -= prob
self.remove_for_transition(xrp, evalnode)
if xrp.resample:
return
xrp = evalnode.xrp
try: # TODO: dont do this here.. dont do cancellign
self.old_to_new_q += math.log(xrp.weight(evalnode.args))
except:
pass # This fails when restoring/keeping, for example
if evalnode.hashval not in self.choices:
raise RException("Choices did not already have this evalnode")
else:
del self.choices[evalnode.hashval]
self.delete_from_db(evalnode.hashval)
def delete_from_db(self, hashval):
if self.db.__contains__(hashval):
self.db.__delitem__(hashval)
elif self.weighted_db.__contains__(hashval):
self.weighted_db.__delitem__(hashval)
def randomKey(self):
if rrandom.random.random() * self.weight() > self.db.weight():
return self.weighted_db.randomKey()
else:
return self.db.randomKey()
def infer(self):
try:
hashval = self.randomKey()
except:
raise RException("Program has no randomness!")
old_p, old_to_new_q, new_p, new_to_old_q = self.reflip(hashval)
p = rrandom.random.random()
if new_p + self.new_to_old_q - old_p - self.old_to_new_q < math.log(p):
self.restore()
else:
self.keep()
self.add_accepted_proposal(hashval)
self.add_made_proposal(hashval)
def reset(self):
self.__init__()
def __str__(self):
string = "EvalNodeTree:"
for evalnode in self.assumes.values():
string += evalnode.str_helper()
for evalnode in self.observes.values():
string += evalnode.str_helper()
for evalnode in self.predicts.values():
string += evalnode.str_helper()
return string
class RandomDB(Engine):
def __init__(self):
#self.db = {}
self.engine_type = 'randomdb'
self.db = RandomChoiceDict()
self.db_noise = {}
self.log = []
# ALWAYS WORKING WITH LOG PROBABILITIES
self.uneval_p = 0
self.eval_p = 0
self.p = 0
self.env = Environment()
self.assumes = {}
self.observes = {}
self.predicts = {}
self.vars = {}
def reset(self):
self.__init__()
def assume(self, varname, expr, id):
self.assumes[id] = (varname, expr)
self.vars[varname] = expr
value = self.evaluate(expr, self.env, reflip = True, stack = [id])
self.env.set(varname, value)
return value
def observe(self, expr, obs_val, id):
if expr.hashval in self.observes:
raise RException('Already observed %s' % str(expr))
self.observes[id] = (expr, obs_val)
# bit of a hack, here, to make it recognize same things as with noisy_expr
self.evaluate(expr, self.env, reflip = False, stack = [id], xrp_force_val = obs_val)
return expr.hashval
def predict(self, expr, id):
self.predicts[id] = expr
return self.evaluate(expr, self.env, True, [id])
def forget(self, id):
self.remove(['obs', id])
assert id in self.observes
del self.observes[id]
return
def insert(self, stack, xrp, value, args, is_obs_noise = False, memorize = True):
stack = tuple(stack)
if self.has(stack):
self.remove(stack)
prob = xrp.prob(value, args)
self.p += prob
xrp.incorporate(value, args)
if is_obs_noise:
self.db_noise[stack] = (xrp, value, args, True)
else:
self.db[stack] = (xrp, value, args, False)
if not is_obs_noise:
self.eval_p += prob # hmmm..
if memorize:
self.log.append(('insert', stack, xrp, value, args, is_obs_noise))
def remove(self, stack, memorize = True):
stack = tuple(stack)
assert self.has(stack)
(xrp, value, args, is_obs_noise) = self.get(stack)
xrp.remove(value, args)
prob = xrp.prob(value, args)
self.p -= prob
if is_obs_noise:
del self.db_noise[stack]
else:
del self.db[stack]
self.uneval_p += prob # previously unindented...
if memorize:
self.log.append(('remove', stack, xrp, value, args, is_obs_noise))
def has(self, stack):
stack = tuple(stack)
return ((stack in self.db) or (stack in self.db_noise))
def get(self, stack):
stack = tuple(stack)
if stack in self.db:
return self.db[stack]
elif stack in self.db_noise:
return self.db_noise[stack]
else:
raise RException('Failed to get stack %s' % str(stack))
def infer(self):
try:
stack = self.random_stack()
except:
return
self.reflip(stack)
def random_stack(self):
key = self.db.randomKey()
return key
def evaluate_recurse(self, subexpr, env, reflip, stack, addition):
newstack = [s for s in stack].extend(addition)
val = self.evaluate(subexpr, env, reflip, newstack)
return val
def binary_op_evaluate(self, expr, env, reflip, stack):
val1 = self.evaluate_recurse(expr.children[0], env, reflip, stack, ['operand0'])
val2 = self.evaluate_recurse(expr.children[1], env, reflip, stack, ['operand1'])
return (val1 , val2)
def children_evaluate(self, expr, env, reflip, stack):
return [self.evaluate_recurse(expr.children[i], env, reflip, stack, ['child' + str(i)]) for i in range(len(expr.children))]
# Draws a sample value (without re-sampling other values) given its parents, and sets it
def evaluate(self, expr, env = None, reflip = False, stack = [], xrp_force_val = None):
if env is None:
env = self.env
if xrp_force_val is not None:
assert expr.type == 'apply'
if expr.type == 'value':
val = expr.val
elif expr.type == 'variable':
var = expr.name
(val, lookup_env) = env.lookup(var)
elif expr.type == 'if':
cond = self.evaluate_recurse(expr.cond, env, reflip, stack , ['cond'])
if cond.bool:
self.unevaluate(stack + ['false'])
val = self.evaluate_recurse(expr.true, env, reflip, stack , ['true'])
else:
self.unevaluate(stack + ['true'])
val = self.evaluate_recurse(expr.false, env, reflip, stack , ['false'])
#elif expr.type == 'switch':
# index = self.evaluate_recurse(expr.index, env, reflip, stack , ['index'])
# assert 0 <= index.num < expr.n
# for i in range(expr.n):
# if i != index.num:
# self.unevaluate(stack + ['child' + str(i)])
# val = self.evaluate_recurse(expr.children[index.num], env, reflip, stack, ['child' + str(index.num)])
elif expr.type == 'let':
# NOTE: this really is a let*
n = len(expr.vars)
assert len(expr.expressions) == n
values = []
new_env = env
for i in range(n): # Bind variables
new_env = new_env.spawn_child()
val = self.evaluate_recurse(expr.expressions[i], new_env, reflip, stack, ['let' + str(i)])
values.append(val)
new_env.set(expr.vars[i], values[i])
if val.type == 'procedure':
val.env = new_env
new_body = expr.body.replace(new_env)
val = self.evaluate_recurse(new_body, new_env, reflip, stack, ['body'])
elif expr.type == 'apply':
n = len(expr.children)
args = [self.evaluate_recurse(expr.children[i], env, reflip, stack, ['arg' + str(i)]) for i in range(n)]
op = self.evaluate_recurse(expr.op, env, reflip, stack , ['operator'])
addition = ','.join([x.str_hash for x in args])
if op.type == 'procedure':
self.unevaluate(stack + ['apply'], addition)
if n != len(op.vars):
raise RException('Procedure should have %d arguments. \nVars were \n%s\n, but children were \n%s.' % (n, op.vars, expr.children))
new_env = op.env.spawn_child()
for i in range(n):
new_env.set(op.vars[i], args[i])
val = self.evaluate_recurse(op.body, new_env, reflip, stack, ['apply', addition])
elif op.type == 'xrp':
self.unevaluate(stack + ['apply'], addition)
if xrp_force_val is not None:
assert not reflip
if self.has(stack):
self.remove(stack)
self.insert(stack, op.xrp, xrp_force_val, args, True)
val = xrp_force_val
else:
substack = stack + ['apply', addition]
if not self.has(substack):
if op.xrp.is_mem_proc():
val = op.xrp.apply_mem(args, self, stack)
else:
val = op.xrp.apply(args)
self.insert(substack, op.xrp, val, args)
else:
if reflip:
self.remove(substack)
if op.xrp.is_mem_proc():
val = op.xrp.apply_mem(args, self, stack)
else:
val = op.xrp.apply(args)
self.insert(substack, op.xrp, val, args)
else:
(xrp, val, dbargs, is_obs_noise) = self.get(substack)
assert not is_obs_noise
else:
raise RException('Must apply either a procedure or xrp. Instead got expression %s' % str(op))
elif expr.type == 'function':
n = len(expr.vars)
new_env = env.spawn_child()
bound = {}
for i in range(n): # Bind variables
bound[expr.vars[i]] = True
procedure_body = expr.body.replace(new_env, bound)
val = Procedure(expr.vars, procedure_body, env)
elif expr.type == '=':
(val1, val2) = self.binary_op_evaluate(expr, env, reflip, stack)
val = val1.__eq__(val2)
elif expr.type == '<':
(val1, val2) = self.binary_op_evaluate(expr, env, reflip, stack)
val = val1.__lt__(val2)
elif expr.type == '>':
(val1, val2) = self.binary_op_evaluate(expr, env, reflip, stack)
val = val1.__gt__(val2)
elif expr.type == '<=':
(val1, val2) = self.binary_op_evaluate(expr, env, reflip, stack)
val = val1.__le__(val2)
elif expr.type == '>=':
(val1, val2) = self.binary_op_evaluate(expr, env, reflip, stack)
val = val1.__ge__(val2)
elif expr.type == '&':
vals = self.children_evaluate(expr, env, reflip, stack)
andval = BoolValue(True)
for x in vals:
andval = andval.__and__(x.bool)
val = andval
elif expr.type == '^':
vals = self.children_evaluate(expr, env, reflip, stack)
xorval = BoolValue(True)
for x in vals:
xorval = xorval.__xor__(x.bool)
val = xorval
elif expr.type == '|':
vals = self.children_evaluate(expr, env, reflip, stack)
orval = BoolValue(False)
for x in vals:
orval = orval.__or__(x.bool)
val = orval
elif expr.type == '~':
negval = self.evaluate_recurse(expr.children[0], env, reflip, stack, ['neg'])
val = negval.__inv__()
elif expr.type == '+':
vals = self.children_evaluate(expr, env, reflip, stack)
sum_val = NatValue(0)
for x in vals:
sum_val = sum_val.__add__(x)
val = sum_val
elif expr.type == '-':
val1 = self.evaluate_recurse(expr.children[0], env, reflip, stack , ['sub0'])
val2 = self.evaluate_recurse(expr.children[1], env, reflip, stack , ['sub1'])
val = val1.__sub__(val2)
elif expr.type == '*':
vals = self.children_evaluate(expr, env, reflip, stack)
prod_val = NatValue(1)
for x in vals:
prod_val = prod_val.__mul__(x)
val = prod_val
elif expr.type == '/':
val1 = self.evaluate_recurse(expr.children[0], env, reflip, stack , ['div0'])
val2 = self.evaluate_recurse(expr.children[1], env, reflip, stack , ['div1'])
val = val1.__div__(val2)
else:
raise RException('Invalid expression type %s' % expr.type)
return val
def unevaluate(self, uneval_stack, args = None):
if args is not None:
args = tuple(args)
to_unevaluate = []
for tuple_stack in self.db:
to_unevaluate.append(tuple_stack)
for tuple_stack in self.db_noise:
to_unevaluate.append(tuple_stack)
to_delete = []
for tuple_stack in to_unevaluate:
stack = list(tuple_stack)
if len(stack) >= len(uneval_stack) and stack[:len(uneval_stack)] == uneval_stack:
if args is None:
to_delete.append(tuple_stack)
else:
assert len(stack) > len(uneval_stack)
if stack[len(uneval_stack)] != args:
to_delete.append(tuple_stack)
for tuple_stack in to_delete:
self.remove(tuple_stack)
def save(self):
self.log = []
self.uneval_p = 0
self.eval_p = 0
def restore(self):
self.log.reverse()
for (type, stack, xrp, value, args, is_obs_noise) in self.log:
if type == 'insert':
self.remove(stack, False)
else:
assert type == 'remove'
self.insert(stack, xrp, value, args, is_obs_noise, False)
def reflip(self, stack):
(xrp, val, args, is_obs_noise) = self.get(stack)
#debug = True
debug = False
old_p = self.p
old_to_new_q = - math.log(len(self.db))
if debug:
print "old_db", self
self.save()
self.remove(stack)
if xrp.is_mem_proc():
new_val = xrp.apply(args, list(stack))
else:
new_val = xrp.apply(args)
self.insert(stack, xrp, new_val, args)
if debug:
print "\nCHANGING ", stack, "\n TO : ", new_val, "\n"
if val == new_val:
return
self.rerun(False)
new_p = self.p
new_to_old_q = self.uneval_p - math.log(len(self.db))
old_to_new_q += self.eval_p
if debug:
print "new db", self, \
"\nq(old -> new) : ", old_to_new_q, \
"q(new -> old) : ", new_to_old_q, \
"p(old) : ", old_p, \
"p(new) : ", new_p, \
'log transition prob : ', new_p + new_to_old_q - old_p - old_to_new_q , "\n"
if old_p * old_to_new_q > 0:
p = rrandom.random.random()
if new_p + new_to_old_q - old_p - old_to_new_q < math.log(p):
self.restore()
if debug:
print 'restore'
self.rerun(False)
if debug:
print "new db", self
print "\n-----------------------------------------\n"
def __str__(self):
string = 'DB with state:'
string += '\n Regular Flips:'
for s in self.db:
string += '\n %s <- %s' % (self.db[s][1].val, s)
string += '\n Observe Flips:'
for s in self.db_noise:
string += '\n %s <- %s' % (self.db_noise[s][1].val, s)
return string
def __contains__(self, stack):
return self.has(self, stack)
def __getitem__(self, stack):
return self.get(self, stack)