def generate(
active_degree: int,
passive_degree: int,
label_count: int,
actives_all_same: bool,
passives_all_same: bool,
flags: ProblemFlags,
count_limit: int = sys.maxsize, # TODO: remove the param or use it
skip_count: int = 0, # TODO: remove the param or use it
) -> List[P]:
alphabet = letter_range(label_count)
# take active_degree labels
# from a pallete of active_label_count
if flags.is_directed_or_rooted:
# rooted/directed: order of configs does not matter,
# except for the frist label in each config. Thus,
# we have additional product() call below.
# e.g. degree = 3, labels = 2, gives us:
# ['AAA', 'AAB', 'ABB', 'BAA', 'BAB', 'BBB']
actives = (
"".join(x)
for x in combinations_with_replacement(alphabet, active_degree -
1))
passives = (
"".join(x)
for x in combinations_with_replacement(alphabet, passive_degree -
1))
actives = ("".join(x) for x in product(alphabet, actives))
passives = ("".join(x) for x in product(alphabet, passives))
else:
# unrooted/undirected: order of configs does not matter
# e.g. degree 3, labels = 2, gives us
# ['AAA', 'AAB', 'ABB', 'BBB']
actives = (
"".join(x)
for x in combinations_with_replacement(alphabet, active_degree))
passives = (
"".join(x)
for x in combinations_with_replacement(alphabet, passive_degree))
if actives_all_same:
actives = (x for x in actives if x[0] * len(x) == x)
if passives_all_same:
passives = (x for x in passives if x[0] * len(x) == x)
active_constraints = (tuple([" ".join(y) for y in x])
for x in tqdm(powerset(actives)) if x)
passive_constraints = (tuple([" ".join(y) for y in x])
for x in tqdm(powerset(passives)) if x)
problem_tuples = (
(a, b) for (a, b) in product(active_constraints, passive_constraints))
return problem_from_constraints(problem_tuples, flags)
def computeIISsBruteForce(self):
"""
DEPRECATED not an efficient way to compute IISs.
If all IIS contain at least one free feature, then safe policies exist.
a brute force way to find all iiss and return their indicies
can be O(2^|unknown features|)
"""
unknownConsPowerset = powerset(self.unknownCons)
feasible = {}
iiss = []
for cons in unknownConsPowerset:
# if any subset is already infeasible, no need to check this set. it's definitely infeasible and not an iis
if len(cons) > 0 and any(not feasible[subset] for subset in combinations(cons, len(cons) - 1)):
feasible[cons] = False
continue
# find if the lp is feasible by posing cons
sol = self.findConstrainedOptPi(cons)
feasible[cons] = sol['feasible']
if len(cons) == 0 and not feasible[cons]:
# no iiss in this case. problem infeasible
return []
# if it is infeasible and all its subsets are feasible (only need to check the subsets with one less element)
# then it's an iis
if not feasible[cons] and all(feasible[subset] for subset in combinations(cons, len(cons) - 1)):
iiss.append(cons)
self.iiss = iiss
def probOfExistanceOfSafePolicies(self, lockedCons, freeCons):
"""
Compute the probability of existence of at least one safe policies.
This considers the changeabilities of all unknown features.
lockedCons, freeCons: The set of locked and free features.
They might be different from the ones confirmed by querying.
These are hypothetical ones just to compute the corresponding prob.
"""
unknownCons = set(self.consIndices) - set(lockedCons) - set(freeCons)
# \EE[policy exists]
expect = 0
allSubsetsOfUnknownCons = powerset(unknownCons)
for freeSubset in allSubsetsOfUnknownCons:
# assume now freeSubset is free and uknownCons \ freeSubset is locked
# compute the prob. that this happens (given lockedCons and freeCons)
prob = reduce(mul, map(lambda _: self.consProbs[_], freeSubset), 1) *\
reduce(mul, map(lambda _: 1 - self.consProbs[_], unknownCons - set(freeSubset)), 1)
# an indicator represents if safe policies exist
safePolicyExist = self.safePolicyExist(freeCons=list(freeCons) +
list(freeSubset))
expect += safePolicyExist * prob
return expect
def computeIISsBruteForce(self):
"""
DEPRECATED not an efficient way to compute IISs.
If all IIS contain at least one free feature, then safe policies exist.
a brute force way to find all iiss and return their indicies
can be O(2^|unknown features|)
"""
unknownConsPowerset = powerset(self.allCons)
feasible = {}
iiss = []
for cons in unknownConsPowerset:
# if any subset is already infeasible, no need to check this set. it's definitely infeasible and not an iis
if len(cons) > 0 and any(not feasible[subset] for subset in combinations(cons, len(cons) - 1)):
feasible[cons] = False
continue
# find if the lp is feasible by posing cons
sol = self.findConstrainedOptPi(cons)
feasible[cons] = sol['feasible']
if len(cons) == 0 and not feasible[cons]:
# no iiss in this case. problem infeasible
return []
# if it is infeasible and all its subsets are feasible (only need to check the subsets with one less element)
# then it's an iis
if not feasible[cons] and all(feasible[subset] for subset in combinations(cons, len(cons) - 1)):
iiss.append(cons)
self.iiss = iiss
def probOfExistanceOfSafePolicies(self, lockedCons, freeCons):
"""
Compute the probability of existence of at least one safe policies.
This considers the changeabilities of all unknown features.
lockedCons, freeCons: The set of locked and free features.
They might be different from the ones confirmed by querying.
These are hypothetical ones just to compute the corresponding prob.
"""
unknownCons = set(self.consIndices) - set(lockedCons) - set(freeCons)
# \EE[policy exists]
expect = 0
allSubsetsOfUnknownCons = powerset(unknownCons)
for freeSubset in allSubsetsOfUnknownCons:
# assume now freeSubset is free and uknownCons \ freeSubset is locked
# compute the prob. that this happens (given lockedCons and freeCons)
prob = reduce(mul, map(lambda _: self.consProbs[_], freeSubset), 1) *\
reduce(mul, map(lambda _: 1 - self.consProbs[_], unknownCons - set(freeSubset)), 1)
# an indicator represents if safe policies exist
safePolicyExist = self.safePolicyExist(freeCons=list(freeCons) + list(freeSubset))
expect += safePolicyExist * prob
return expect
def getProbOfExistenceOfSafePolicies(self, lockedCons, freeCons):
"""
Compute the probability that safe policies exist using dominating policies (the best way?)
lockedCons, freeCons: The set of locked and free features.
They might be different from the ones confirmed by querying.
These are hypothetical ones just to compute the corresponding prob.
"""
result = 0
def pf(con):
if con in lockedCons: return 0
elif con in freeCons: return 1
else: return self.consProbs[con]
assert hasattr(self, 'domPiFeats')
# two ways to compute the probs. either 2^|relFeats| or 2^|domPis|
# so see which one is smaller
if len(self.relFeats) < len(self.domPiFeats):
allSubsetsOfRelFeats = powerset(self.relFeats)
for freeSubset in allSubsetsOfRelFeats:
if self.safePolicyExist(freeCons = list(freeSubset) + list(self.knownFreeCons)):
prob = self.probFeatsBeingFree(freeSubset) * self.probFeatsBeingLocked(set(self.relFeats) - set(freeSubset))
result += prob
else:
for k in range(1, len(self.domPiFeats) + 1):
sign = 1 if k % 2 == 1 else -1
for domPiFeatsSubset in combinations(self.domPiFeats, k):
domPiFeatsSubsetLists = map(lambda _: list(_), domPiFeatsSubset)
unionOfFeats = set(sum(domPiFeatsSubsetLists, []))
result += sign * reduce(mul, map(pf, unionOfFeats), 1)
return result
def findRelevantFeaturesBruteForce(self):
"""
a method simply to measure the time needed to compute all dominating policies
"""
allConsPowerset = set(powerset(self.unknownCons))
for subsetsToConsider in allConsPowerset:
self.findConstrainedOptPi(subsetsToConsider)
def _fifteens(self):
"""Compute the score for fifteens"""
fifteen_score = 0
pset = powerset(self.all_five)
for entry in pset:
if self._sum_cards(entry) == 15:
fifteen_score += 2
return fifteen_score
def knapsack(loot, weight_limit):
all_combo = powerset(loot)
best_value = None
for combo in all_combo:
combo_weight = sum([item.weight for item in combo])
combo_value = sum([item.value for item in combo])
if combo_weight <= weight_limit:
if not best_value or best_value < combo_value:
best_value = combo_value
if not best_value:
print("knapsack couldn't fit any items")
return best_value
def _runs(self):
"""Compute the score for a run of three, four, or five cards"""
# Used for sorting subsets
def get_card_value(card):
return card.value
runs = []
run_score = 0
pset = powerset(self.all_five)
for entry in pset:
sub_run = False
if len(entry) < 3:
continue
sorted_entry = sorted(entry, key=get_card_value)
current_val = sorted_entry[0].value
count = 1
for card in sorted_entry[1:]:
if card.value == (current_val + 1):
count += 1
current_val = card.value
else:
count = 0
if count == 0:
continue
if count >= 3:
runs.extend([sorted_entry])
runs_final = []
for run1 in runs:
sublist = False
for run2 in runs:
if self._sublist(run1, run2):
sublist = True
if not sublist:
runs_final.extend([run1])
for run in runs_final:
run_score += len(run)
return run_score
def divisors(n):
"""A set of the proper divisors of n for any positive integer n.
>>> divisors(1)
{1}
>>> divisors(17.0)
{1}
>>> divisors(4) == {1, 2}
True
>>> divisors(49) == {1, 7}
True
>>> divisors(120) == {1, 2, 3, 4, 5, 6, 8, 10, 12, 15, 20, 24, 30, 40, 60}
True
>>> len(divisors(360))
23
>>> len(divisors(2520))
47
>>> divisors(0)
Traceback (most recent call last):
...
ValueError: n must be positive
>>> divisors(2**0.5)
Traceback (most recent call last):
...
ValueError: n must be an integer
"""
from functools import reduce
from math import floor
from util import powerset, prime_factorization
if floor(n) != n:
raise ValueError("n must be an integer")
n = floor(n)
if not n >= 1:
raise ValueError("n must be positive")
pf = prime_factorization(n)
factor_sets = powerset(pf)
product = lambda factors: reduce(lambda x,y: x*y, factors, 1)
divisors = {product(factors) for factors in factor_sets}
if n != 1:
divisors.discard(n)
return divisors
def choice(self, contracts):
print "choosing from: ",
for c in contracts:
print c.id,
print ""
# Generate power set of possible contracts
allContractSets = util.powerset(contracts)
maxU = 0
maxContractSet = []
for contractSet in allContractSets:
u = self.utility(self, contractSet)
if u > maxU:
maxU = u
maxContractSet = contractSet
return maxContractSet
def findRelevantFeaturesAndDomPis(self):
"""
Incrementally add dominating policies to a set
DomPolicies algorithm in the IJCAI paper
"""
beta = [] # rules to keep
dominatingPolicies = {}
allCons = set()
allConsPowerset = set(powerset(allCons))
subsetsConsidered = []
# iterate until no more dominating policies are found
while True:
subsetsToConsider = allConsPowerset.difference(subsetsConsidered)
if len(subsetsToConsider) == 0: break
# find the subset with the smallest size
activeCons = min(subsetsToConsider, key=lambda _: len(_))
#if config.DEBUG: print 'activeCons', activeCons
subsetsConsidered.append(activeCons)
skipThisCons = False
for enf, relax in beta:
if enf.issubset(activeCons) and len(relax.intersection(activeCons)) == 0:
# this subset can be ignored
skipThisCons = True
break
if skipThisCons:
continue
sol = self.findConstrainedOptPi(activeCons)
if sol['feasible']:
x = sol['pi']
if config.DEBUG:
printOccSA(x)
print self.computeValue(x)
dominatingPolicies[activeCons] = x
# check violated constraints
violatedCons = self.findViolatedConstraints(x)
if config.DEBUG: print 'x violates', violatedCons
else:
# infeasible
violatedCons = ()
if config.DEBUG: print 'infeasible'
# beta records that we would not enforce activeCons and relax occupiedFeats in the future
beta.append((set(activeCons), set(violatedCons)))
allCons.update(violatedCons)
allConsPowerset = set(powerset(allCons))
domPis = []
for pi in dominatingPolicies.values():
if pi not in domPis: domPis.append(pi)
if config.DEBUG: print 'rel cons', allCons, 'num of domPis', len(domPis)
return allCons, domPis
def findRelevantFeaturesBruteForce(self):
allConsPowerset = set(powerset(self.allCons))
for subsetsToConsider in allConsPowerset:
self.findConstrainedOptPi(subsetsToConsider)
def computeOptQueries(self):
"""
f(\phi_l, \phi_f) =
0, if safePolicyExist(\phi_f) or self.safePolicyNotExist(\phi_l)
min_\phi p_f(\phi) f(\phi_l, \phi_f + {\phi}) + (1 - p_f(\phi)) f(\phi_l + {\phi}, \phi_f), o.w.
Boundaries condition:
\phi_l is not a superset of any iis, \phi_f is not a superset of rel feats of any dom pi, otherwise 0 for sure
"""
consPowerset = list(powerset(self.relFeats))
# free/locked cons that are not supersets of elements on their boundaries
admissibleFreeCons = []
admissibleLockedCons = []
# the set of (lockedCons, freeCons) to evaluate the optimal queries
# it's the cross product of the two sets above, excluding free and locked cons that share elements
admissibleCons = []
for lockedCons in consPowerset:
if self.safePolicyNotExist(lockedCons=lockedCons):
if not any(set(lockedCons).issuperset(lockedB) for lockedB in self.lockedBoundary):
self.lockedBoundary.append(lockedCons)
else: admissibleLockedCons.append(lockedCons)
for freeCons in consPowerset:
if self.safePolicyExist(freeCons=freeCons):
if not any(set(freeCons).issuperset(freeB) for freeB in self.freeBoundary):
self.freeBoundary.append(freeCons)
else: admissibleFreeCons.append(freeCons)
if config.VERBOSE:
print 'locked', self.lockedBoundary
print 'free', self.freeBoundary
for lockedCons in admissibleLockedCons:
for freeCons in admissibleFreeCons:
# any cons should not be known to be both free and locked
if set(lockedCons).isdisjoint(set(freeCons)):
admissibleCons.append((lockedCons, freeCons))
readyToEvaluate = lambda l, f: all(self.getQueryAndValue(l, set(f).union({con})) != None \
and self.getQueryAndValue(set(l).union({con}), f) != None \
for con in set(self.relFeats) - set(l) - set(f))
# keep the sets of cons that are ready to evaluate in the next iteration
readyToEvalSet = []
for (lockedCons, freeCons) in admissibleCons:
if readyToEvaluate(lockedCons, freeCons): readyToEvalSet.append((lockedCons, freeCons))
# keep fill out the values of optQs within boundary
# whenever filled out
while len(readyToEvalSet) > 0:
if config.VERBOSE: print len(readyToEvalSet), 'need to be evaluated'
(lockedCons, freeCons) = readyToEvalSet.pop()
unknownCons = set(self.relFeats) - set(lockedCons) - set(freeCons)
minNums = [(con,\
self.consProbs[con] * self.getQueryAndValue(lockedCons, set(freeCons).union({con}))[1]\
+ (1 - self.consProbs[con]) * self.getQueryAndValue(set(lockedCons).union({con}), freeCons)[1]\
+ 1) # count con in
for con in unknownCons]
# pick the tuple that has the minimum obj value after querying
self.setQueryAndValue(lockedCons, freeCons, min(minNums, key=lambda _: _[1]))
# add neighbors that ready to evaluate to readToEvalSet
readyToEvalSet += filter(lambda (l, f): self.getQueryAndValue(l, f) == None and readyToEvaluate(l, f),\
[(set(lockedCons) - {cons}, freeCons) for cons in lockedCons] +\
[(lockedCons, set(freeCons) - {cons}) for cons in freeCons])
def computeOptimalQuery(self, knownLockedCons, knownFreeCons, unknownCons,
psi):
"""
recursively compute the optimal query, return the value after query
"""
# the key used for optQueryAndValueDict
# use frozenset here because the order of features doesn't matter
key = (frozenset(knownLockedCons), frozenset(knownFreeCons),
frozenset(unknownCons), tuple(psi))
if key in self.optQueryAndValueDict.keys():
return self.optQueryAndValueDict[key]
rewardSupports = self.computeConsistentRewardIndices(psi)
self.imaginedMDP.updatePsi(psi)
# compute the current safe policy
if key in self.currentOptPiValueDict.keys():
currentSafelyOptValue = self.currentOptPiValueDict[key]
else:
currentSafelyOptValue = self.findConstrainedOptPi(
activeCons=list(unknownCons) + list(knownLockedCons),
addKnownLockedCons=False,
mdp=self.imaginedMDP)['obj']
# feature queries
if len(unknownCons) > 0:
consQueryValues = {
('F', con): self.consProbs[con] * self.computeOptimalQuery(
knownLockedCons, knownFreeCons + [con],
set(unknownCons) - {con}, psi)[1] +
(1 - self.consProbs[con]) * self.computeOptimalQuery(
knownLockedCons + [con], knownFreeCons,
set(unknownCons) - {con}, psi)[1] - self.costOfQuery
for con in unknownCons
}
else:
consQueryValues = {}
# reward queries
psiOfSet = lambda rSet: sum(psi[_] for _ in rSet)
if len(rewardSupports) > 1:
rewardQueryValues = {
('R', rSet): psiOfSet(rSet) * self.computeOptimalQuery(
knownLockedCons, knownFreeCons, unknownCons,
computePosteriorBelief(psi, consistentRewards=rSet))[1] +
(1 - psiOfSet(rSet)) * self.computeOptimalQuery(
knownLockedCons, knownFreeCons, unknownCons,
computePosteriorBelief(psi, inconsistentRewards=rSet))[1] -
self.costOfQuery
for rSet in powerset(
rewardSupports, minimum=1, maximum=len(rewardSupports) - 1)
}
else:
rewardQueryValues = {}
queryAndValues = consQueryValues.copy()
queryAndValues.update(rewardQueryValues)
# also, there's an option to not pose a query
queryAndValues[None] = currentSafelyOptValue
optQueryAndValue = max(queryAndValues.items(), key=lambda _: _[1])
self.optQueryAndValueDict[key] = optQueryAndValue
return optQueryAndValue
def findDomPi(self):
"""
(re)compute all dominating policies given reward and safety uncertainty
and then sample one
stored in self.dompis = [(dompi, weighted_prob)]
"""
domPisData = []
allDomPis = []
priorPi = self.computeCurrentSafelyOptPi()
consistentRewardIndices = self.computeConsistentRewardIndices(
self.mdp.psi)
for rIndices in powerset(consistentRewardIndices,
minimum=1,
maximum=self.sizeOfRewards):
rewardPositiveMDP = copy.deepcopy(self.mdp)
rewardPositiveMDP.updatePsi(
computePosteriorBelief(self.mdp.psi,
consistentRewards=rIndices))
sumOfPsi = sum(self.mdp.psi[_] for _ in rIndices)
rewardPositiveConsAgent = ConsQueryAgent(
rewardPositiveMDP,
self.consStates,
self.goalCons,
self.consProbs,
knownFreeCons=self.knownFreeCons,
knownLockedCons=self.knownLockedCons)
_, domPis = rewardPositiveConsAgent.findRelevantFeaturesAndDomPis()
for domPi in domPis:
relFeats = rewardPositiveConsAgent.findViolatedConstraints(
domPi)
# we are going to query about rIndices and relFeatures
# we regard them as batch queries and compute the possible responses
safeProb = numpy.prod(
[self.consProbs[feat] for feat in relFeats])
rPositiveValue = rewardPositiveConsAgent.computeValue(domPi)
# priorPi is feasible under relFeats since priorPi is safer (before querying)
priorValue = rewardPositiveConsAgent.computeValue(priorPi)
# 1 <= len(rIndices) <= sizeOfRewards
rewardQueryNeeded = (len(rIndices) <
len(consistentRewardIndices))
# at least (relFeats) feature queries and 1 reward-set query are needed
weightedValue = safeProb * sumOfPsi * (
rPositiveValue - priorValue - self.costOfQuery *
(len(relFeats) + rewardQueryNeeded))
if domPi not in allDomPis:
allDomPis.append(domPi)
if weightedValue > 0:
# only add dom pi info when it's beneficial to query about this
domPisData.append(
self.DomPiData(pi=domPi,
weightedValue=weightedValue,
optimizedRewards=rIndices,
violatedCons=relFeats))
if self.domPiNum is None:
self.domPiNum = len(allDomPis)
if len(domPisData) > 0:
self.objectDomPiData = max(domPisData,
key=lambda datum: datum.weightedValue)
else:
self.objectDomPiData = None
if config.VERBOSE: print 'chosen dom pi', self.objectDomPiData
def _satisfactory_placement_generator(self):
# generates all satisfactory mine placements
# note the use of of powerset, a function (not method)
# defined in this using the recipe from itertools package
return filter(self._is_satisfactory_placement, powerset(self.perimiter))
def findRelevantFeaturesAndDomPis(self):
"""
Incrementally add dominating policies to a set
DomPolicies algorithm in the IJCAI paper
"""
beta = [] # rules to keep
dominatingPolicies = {}
allCons = set()
allConsPowerset = set(powerset(allCons))
subsetsConsidered = []
# iterate until no more dominating policies are found
while True:
subsetsToConsider = allConsPowerset.difference(subsetsConsidered)
if len(subsetsToConsider) == 0: break
# find the subset with the smallest size
activeCons = min(subsetsToConsider, key=lambda _: len(_))
#if config.DEBUG: print 'activeCons', activeCons
subsetsConsidered.append(activeCons)
skipThisCons = False
for enf, relax in beta:
if enf.issubset(activeCons) and len(
relax.intersection(activeCons)) == 0:
# this subset can be ignored
skipThisCons = True
break
if skipThisCons:
continue
sol = self.findConstrainedOptPi(activeCons)
if sol['feasible']:
x = sol['pi']
if config.DEBUG:
printOccSA(x)
print self.computeValue(x)
dominatingPolicies[activeCons] = x
# check violated constraints
violatedCons = self.findViolatedConstraints(x)
if config.DEBUG: print 'x violates', violatedCons
else:
# infeasible
violatedCons = ()
if config.DEBUG: print 'infeasible'
# beta records that we would not enforce activeCons and relax occupiedFeats in the future
beta.append((set(activeCons), set(violatedCons)))
allCons.update(violatedCons)
allConsPowerset = set(powerset(allCons))
domPis = []
for pi in dominatingPolicies.values():
if pi not in domPis: domPis.append(pi)
if config.DEBUG:
print 'rel cons', allCons, 'num of domPis', len(domPis)
return allCons, domPis
def best(patterns, extended):
# Enumberate daytimes
dts = collections.defaultdict(set)
all = []
for p in patterns:
all.extend(p[0].each())
for p in all:
for dt in p.getDayTimesRaw():
dts[dt.data()].add(p)
# For each daytime, iterate patterns to find termweeks
dt_tw = {}
dt_tw_sz = {}
for (dt, ps) in dts.iteritems():
tws = collections.defaultdict(set)
for p in ps:
for (term, week) in p.getTermWeeks().each():
tws[term].add(week)
dt_tw[dt] = tws
dt_tw_sz[dt] = reduce(lambda tot, item: tot + len(item),
tws.itervalues(), 0)
# restrict to at most max_trials (longest)
dt_use = set()
dt_candidates = dt_tw.keys()
for i in range(0, max_trials):
if len(dt_candidates) == 0:
break
use = max(dt_candidates, key=lambda k: dt_tw_sz[k])
dt_candidates.remove(use)
dt_use.add(use)
# find longest range of each, using 1-8,9-16,17-24 type ranges to allow term overlap
dt_longest = {}
for dt in dt_use:
# build termy week numbers (1-24)
week_nums = set()
for (term, weeks) in dt_tw[dt].iteritems():
for week in filter(lambda x: x > 0 and x < 9, weeks):
week_nums.add(term * 8 + week)
ranges = sorted(util.ranges(week_nums),
key=lambda x: x[1],
reverse=True)
if len(ranges) == 0:
dt_longest[dt] = set()
else:
dt_longest[dt] = set(
range(ranges[0][0], ranges[0][0] + ranges[0][1]))
# permute through including and excluding date ranges to see which gives best coverage (EXPONENTIAL!)
best_score = None
best = None
for dts in util.powerset(dt_use):
if len(dts) == 0:
continue
all = set(range(1, 25))
for dt in dts:
all &= dt_longest[dt]
score = len(all) * len(dts)
if best_score == None or score > best_score:
best_score = score
best = dts
# Generate pattern
if best is None:
logger.error("No common in %s" % all)
return None
p = patternatom.PatternAtom(False)
for b in best:
p.addDayTimeRange(b[0], b[1][0], b[1][1], b[2][0], b[2][1])
p.setAllYear()
# Extend to include out-of-term dates, where required
if extended:
for q in patterns:
for qa in q[0].blast():
p.expand_back_to(qa)
p.expand_forward_to(qa)
return p
def execute(args):
print 'Starting the artificial neural network'
if len(args) < 2:
usage()
sys.exit()
###############################################################################
# Data
# names feature labels
# y shuffled names
# x features that correspond to shuffled names
names, y, x = parse(args[1])
x = clean(names, x)
usePowerset = args[0]
# Build features to include in test
features = args[2:]
if len(features) == 0:
features = names
# print 'Selected features:', features
# Build all subsets of features, if requested
if usePowerset.lower() == 'true':
combos = powerset(features)
else:
combos = [features]
# map from feature set, to map of correct counts for each person
feature_performance = {}
highest_correct = 0
best_combo = {}
for c in combos:
if len(c) == 0:
continue
print 'Attempting feature set:', c
x_selected = selectFeatures(copy.copy(names), c, x)
# Split into testing and traiing data
x_train, x_test, y_train, y_test = train_test_split(x_selected, y,
test_size=0.2,
random_state=0)
###############################################################################
# Models
logistic = linear_model.LogisticRegression(C=L_REGULARIZATION)
rbm = BernoulliRBM(random_state=0, verbose=True, learning_rate=N_LEARNING_RATE, n_iter=N_ITER, n_components=N_COMPONENTS)
# Note: attempted StandardScaler, MinMaxScaler, MaxAbsScaler, without strong results
# Not needed, since data is scaled to the [0-1] range by clean()
classifier = Pipeline(steps=[('rbm', rbm),('logistic', logistic)])
# ###############################################################################
# Training
print 'Training the classifier...'
# Training RBM-Logistic Pipeline
classifier.fit(x_train,y_train)
correct = 0
label_counts = defaultdict(int)
for i in range(len(x_test)):
test = x_test[i]
if len(test) == 1:
test = test.reshape(-1, 1)
else:
test = [test]
predicted = classifier.predict(test)
if predicted == y_test[i]:
correct += 1
label_counts[predicted[0]] += 1
if correct >= highest_correct:
highest_correct = correct
best_combo = c
feature_performance[str(c)] = {'predictions':label_counts,'expected':Counter(y_test)}
###############################################################################
# Evaluation
# evaluate(classifier, x_test, y_test)
summary = feature_performance[str(best_combo)]
print 'Accuracy:\t\t\t', highest_correct, 'correct gives', (highest_correct * 1.0/len(y_test)), 'compared to guessing', (1.0/len(summary['expected']))
print 'Best feature set:\t\t', best_combo
print 'Identified %d out of %d labels'%(len(summary['predictions']),len(summary['expected']))
for p in summary['predictions']:
pred = summary['predictions'][p]
tot = summary['expected'][p]
print '\t %s \t\t %d\t of %d \t (%f)'%(p, pred, tot, pred * 1.0/tot)
def _gcjut_rec(loop, t, conj_existing_states=False):
# Base case
if type(t) == Var or type(t) == Const:
return [Join(loop, t)]
out = []
# Recursively call function on subterms
joins = [_gcjut_rec(loop, st) for st in t.terms]
for j_comb in product(*joins):
# For this particular combination of joins, obtain a merged join
merged_join = merge(loop, t.op, j_comb)
vprint(P_JOIN_GEN, "Join: merged these joins:")
for join in j_comb:
vprint(P_JOIN_GEN, "Join:", join)
out.append(merged_join) # Case when merged join is not a new auxillary
vprint(P_JOIN_GEN, "Join: candidate join (merged) =\n", merged_join)
if not merged_join.term.state_free("SV"):
continue
# Find all constants and obtain a mapping to their locations
const_indv = _get_const_indv(merged_join.term)
if not const_indv:
return out
for const in const_indv.keys():
vprint(
P_JOIN_GEN,
"Join: const %s appears in locations %s within %s)" %
(str(const), str(const_indv[const]), str(merged_join.term)))
for ind_set in powerset(const_indv[const]):
if not ind_set:
continue
rem_set = const_indv[const][:]
auxjn = Join(merged_join.loop, merged_join.term)
k = auxjn.loop.get_num_states()
# Conjecture that this particular choice of indices corresponds
# to locations of an auxillary state variable
for ind in ind_set:
auxjn.term.set_term_at(ind, Var("RSV", "s", k + 1))
rem_set.remove(ind)
# Unfold right variables in term to obtain definition for auxillary
auxterm = auxjn.term.rename("RSV", "SV").apply_subst(
merged_join.loop.get_full_state_subst())
# For all remaining indices, conjecture that some of them point to
# existing state variables (if conj_existing_states is True)
for state_assgn in product(
*[list(range(loop.get_num_states() + 1)) for _ in range(len(rem_set))]) \
if conj_existing_states else [[0] * len(rem_set)]:
auxjn_v = deepcopy(auxjn)
auxterm_v = deepcopy(auxterm)
for i in range(len(rem_set)):
if state_assgn[i] != 0:
auxterm_v.set_term_at(
rem_set[i], Var("SV", "s", state_assgn[i]))
# Add the auxillary variable and set the join to be the auxillary
# Note: the auxillary variable could already exist among the states,
# in which case, r is an index to the existing state
r = auxjn_v.loop.add_state(const, auxterm_v, k)
auxjn_v.term = Var("RSV", "s", r + 1)
out.append(auxjn_v)
vprint(P_JOIN_GEN, "Join: new auxillary variable:")
vprint(
P_JOIN_GEN,
"Join: %s = %s" % (str(auxjn_v.term), str(auxterm_v)))
vprint(P_JOIN_GEN,
"Join: candidate join (with auxillaries) =\n",
str(auxjn_v))
return out
def execute(args):
print 'Starting the artificial neural network'
if len(args) < 2:
usage()
sys.exit()
###############################################################################
# Data
# names feature labels
# y shuffled names
# x features that correspond to shuffled names
names, y, x = parse(args[1])
x = clean(names, x)
usePowerset = args[0]
# Build features to include in test
features = args[2:]
if len(features) == 0:
features = names
# print 'Selected features:', features
# Build all subsets of features, if requested
if usePowerset.lower() == 'true':
combos = powerset(features)
else:
combos = [features]
# map from feature set, to map of correct counts for each person
feature_performance = {}
highest_correct = 0
best_combo = {}
for c in combos:
if len(c) == 0:
continue
print 'Attempting feature set:', c
x_selected = selectFeatures(copy.copy(names), c, x)
# Split into testing and traiing data
x_train, x_test, y_train, y_test = train_test_split(x_selected,
y,
test_size=0.2,
random_state=0)
###############################################################################
# Models
logistic = linear_model.LogisticRegression(C=L_REGULARIZATION)
rbm = BernoulliRBM(random_state=0,
verbose=True,
learning_rate=N_LEARNING_RATE,
n_iter=N_ITER,
n_components=N_COMPONENTS)
# Note: attempted StandardScaler, MinMaxScaler, MaxAbsScaler, without strong results
# Not needed, since data is scaled to the [0-1] range by clean()
classifier = Pipeline(steps=[('rbm', rbm), ('logistic', logistic)])
# ###############################################################################
# Training
print 'Training the classifier...'
# Training RBM-Logistic Pipeline
classifier.fit(x_train, y_train)
correct = 0
label_counts = defaultdict(int)
for i in range(len(x_test)):
test = x_test[i]
if len(test) == 1:
test = test.reshape(-1, 1)
else:
test = [test]
predicted = classifier.predict(test)
if predicted == y_test[i]:
correct += 1
label_counts[predicted[0]] += 1
if correct >= highest_correct:
highest_correct = correct
best_combo = c
feature_performance[str(c)] = {
'predictions': label_counts,
'expected': Counter(y_test)
}
###############################################################################
# Evaluation
# evaluate(classifier, x_test, y_test)
summary = feature_performance[str(best_combo)]
print 'Accuracy:\t\t\t', highest_correct, 'correct gives', (
highest_correct * 1.0 /
len(y_test)), 'compared to guessing', (1.0 / len(summary['expected']))
print 'Best feature set:\t\t', best_combo
print 'Identified %d out of %d labels' % (len(
summary['predictions']), len(summary['expected']))
for p in summary['predictions']:
pred = summary['predictions'][p]
tot = summary['expected'][p]
print '\t %s \t\t %d\t of %d \t (%f)' % (p, pred, tot,
pred * 1.0 / tot)
def computeOptQueries(self):
"""
f(\phi_l, \phi_f) =
0, if safePolicyExist(\phi_f) or self.safePolicyNotExist(\phi_l)
min_\phi p_f(\phi) f(\phi_l, \phi_f + {\phi}) + (1 - p_f(\phi)) f(\phi_l + {\phi}, \phi_f), o.w.
Boundaries condition:
\phi_l is not a superset of any iis, \phi_f is not a superset of rel feats of any dom pi, otherwise 0 for sure
"""
consPowerset = list(powerset(self.relFeats))
# free/locked cons that are not supersets of elements on their boundaries
# that is, we can't determine if safe policy exsits if these elements are known to be locked/free
admissibleFreeCons = []
admissibleLockedCons = []
# the set of (lockedCons, freeCons) to evaluate the optimal queries
# it's the cross product of the two sets above, excluding free and locked cons that share elements
admissibleCons = []
for lockedCons in consPowerset:
if self.safePolicyNotExist(lockedCons=lockedCons):
# make sure not elements in boundary is a superset of another element
if not any(
set(lockedCons).issuperset(lockedB)
for lockedB in self.lockedBoundary):
self.lockedBoundary.append(lockedCons)
else:
admissibleLockedCons.append(lockedCons)
for freeCons in consPowerset:
if self.safePolicyExist(freeCons=freeCons):
# similarly
if not any(
set(freeCons).issuperset(freeB)
for freeB in self.freeBoundary):
self.freeBoundary.append(freeCons)
else:
admissibleFreeCons.append(freeCons)
if config.DEBUG:
print 'locked', self.lockedBoundary
print 'free', self.freeBoundary
for lockedCons in admissibleLockedCons:
for freeCons in admissibleFreeCons:
# any cons should not be known to be both free and locked
if set(lockedCons).isdisjoint(set(freeCons)):
admissibleCons.append((lockedCons, freeCons))
# make sure all terms on the RHS (of def of f above) are evaluated
readyToEvaluate = lambda l, f: all(self.getQueryAndValue(l, set(f).union({con})) != None \
and self.getQueryAndValue(set(l).union({con}), f) != None \
for con in set(self.relFeats) - set(l) - set(f))
# keep the sets of cons that are ready to evaluate in the next iteration
readyToEvalSet = []
for (lockedCons, freeCons) in admissibleCons:
if readyToEvaluate(lockedCons, freeCons):
readyToEvalSet.append((lockedCons, freeCons))
# keep fill out the values of optQs within boundary
# whenever filled out
while len(readyToEvalSet) > 0:
if config.DEBUG: print len(readyToEvalSet), 'need to be evaluated'
(lockedCons, freeCons) = readyToEvalSet.pop()
unknownCons = set(self.relFeats) - set(lockedCons) - set(freeCons)
# evaluate all candidate con and compute their minimum number of queries
minNums = [(con,
self.consProbs[con] * self.getQueryAndValue(lockedCons, set(freeCons).union({con}))[1]\
+ (1 - self.consProbs[con]) * self.getQueryAndValue(set(lockedCons).union({con}), freeCons)[1]\
+ 1)
for con in unknownCons]
# pick the tuple that has the minimum obj value after querying
self.setQueryAndValue(lockedCons, freeCons, minNums)
# add neighbors that ready to evaluate to readToEvalSet
readyToEvalSet += filter(lambda (l, f): self.getQueryAndValue(l, f) == None and readyToEvaluate(l, f),
[(set(lockedCons) - {cons}, freeCons) for cons in lockedCons] +\
[(lockedCons, set(freeCons) - {cons}) for cons in freeCons])
def findRelevantFeaturesBruteForce(self):
allConsPowerset = set(powerset(self.unknownCons))
for subsetsToConsider in allConsPowerset:
self.findConstrainedOptPi(subsetsToConsider)
def findRelevantFeaturesAndDomPis(self):
"""
Incrementally add dominating policies to a set
DomPolicies algorithm in the IJCAI paper
earlyStop: stop within this time and return whatever dompis found
"""
beta = [] # rules to keep
dominatingPolicies = {}
allCons = set()
allConsPowerset = set(powerset(allCons))
subsetsConsidered = []
if config.earlyStop is None:
# never stop before finding all dom pis
terminateCond = lambda: False
else:
startTime = time.time()
terminateCond = lambda: time.time() - startTime >= config.earlyStop
# iterate until no more dominating policies are found
while not terminateCond():
subsetsToConsider = allConsPowerset.difference(subsetsConsidered)
if len(subsetsToConsider) == 0: break
# find the subset with the smallest size
activeCons = min(subsetsToConsider, key=lambda _: len(_))
if config.DEBUG: print 'activeCons', activeCons
subsetsConsidered.append(activeCons)
skipThisCons = False
for enf, relax in beta:
if enf.issubset(activeCons) and len(
relax.intersection(activeCons)) == 0:
# this subset can be ignored
skipThisCons = True
if config.DEBUG: print 'dominated'
break
if skipThisCons:
continue
# it will enforce activeCons and known locked features (inside)
sol = self.findConstrainedOptPi(activeCons)
if sol['feasible']:
x = sol['pi']
if config.DEBUG:
printOccSA(x)
print self.computeValue(x)
dominatingPolicies[activeCons] = x
# check violated constraints
violatedCons = self.findViolatedConstraints(x)
if config.DEBUG: print 'this policy violates', violatedCons
else:
# infeasible
violatedCons = ()
if config.DEBUG: print 'infeasible'
# beta records that we would not enforce activeCons and relax occupiedFeats in the future
beta.append((set(activeCons), set(violatedCons)))
allCons.update(violatedCons)
allConsPowerset = set(powerset(allCons))
domPis = []
for pi in dominatingPolicies.values():
if pi not in domPis: domPis.append(pi)
# make sure returned values are lists
allCons = list(allCons)
if config.DEBUG:
print 'rel cons', allCons, 'num of domPis', len(domPis)
return allCons, domPis
return raii_tmpdir()
def readdir_inode(dir):
cmd = base_cmdline + [pjoin(basename, 'test', 'readdir_inode'), dir]
with subprocess.Popen(cmd, stdout=subprocess.PIPE,
universal_newlines=True) as proc:
lines = proc.communicate()[0].splitlines()
lines.sort()
return lines
@pytest.mark.parametrize(
"cmdline_builder",
(invoke_directly, invoke_mount_fuse, invoke_mount_fuse_drop_privileges))
@pytest.mark.parametrize("options", powerset(options))
@pytest.mark.parametrize("name", ('hello', 'hello_ll'))
def test_hello(tmpdir, name, options, cmdline_builder, output_checker):
mnt_dir = str(tmpdir)
mount_process = subprocess.Popen(cmdline_builder(mnt_dir, name, options),
stdout=output_checker.fd,
stderr=output_checker.fd)
try:
wait_for_mount(mount_process, mnt_dir)
assert os.listdir(mnt_dir) == ['hello']
filename = pjoin(mnt_dir, 'hello')
with open(filename, 'r') as fh:
assert fh.read() == 'Hello World!\n'
with pytest.raises(IOError) as exc_info:
open(filename, 'r+')
assert exc_info.value.errno == errno.EACCES
def best(patterns,extended):
# Enumberate daytimes
dts = collections.defaultdict(set)
all = []
for p in patterns:
all.extend(p[0].each())
for p in all:
for dt in p.getDayTimesRaw():
dts[dt.data()].add(p)
# For each daytime, iterate patterns to find termweeks
dt_tw = {}
dt_tw_sz = {}
for (dt,ps) in dts.iteritems():
tws = collections.defaultdict(set)
for p in ps:
for (term,week) in p.getTermWeeks().each():
tws[term].add(week)
dt_tw[dt] = tws
dt_tw_sz[dt] = reduce(lambda tot,item: tot+len(item),tws.itervalues(),0)
# restrict to at most max_trials (longest)
dt_use = set()
dt_candidates = dt_tw.keys()
for i in range(0,max_trials):
if len(dt_candidates) == 0:
break
use = max(dt_candidates,key = lambda k: dt_tw_sz[k])
dt_candidates.remove(use)
dt_use.add(use)
# find longest range of each, using 1-8,9-16,17-24 type ranges to allow term overlap
dt_longest = {}
for dt in dt_use:
# build termy week numbers (1-24)
week_nums = set()
for (term,weeks) in dt_tw[dt].iteritems():
for week in filter(lambda x: x>0 and x<9,weeks):
week_nums.add(term*8+week)
ranges = sorted(util.ranges(week_nums),key = lambda x: x[1],reverse = True)
if len(ranges) == 0:
dt_longest[dt] = set()
else:
dt_longest[dt] = set(range(ranges[0][0],ranges[0][0]+ranges[0][1]))
# permute through including and excluding date ranges to see which gives best coverage (EXPONENTIAL!)
best_score = None
best = None
for dts in util.powerset(dt_use):
if len(dts) == 0:
continue
all = set(range(1,25))
for dt in dts:
all &= dt_longest[dt]
score = len(all) * len(dts)
if best_score == None or score > best_score:
best_score = score
best = dts
# Generate pattern
if best is None:
logger.error("No common in %s" % all)
return None
p = patternatom.PatternAtom(False)
for b in best:
p.addDayTimeRange(b[0],b[1][0],b[1][1],b[2][0],b[2][1])
p.setAllYear()
# Extend to include out-of-term dates, where required
if extended:
for q in patterns:
for qa in q[0].blast():
p.expand_back_to(qa)
p.expand_forward_to(qa)
return p
