def findTotalMisWithMCSAT(self, finalout):
for key, value in self.weightsDic.items():
value['nii'] = 0
if isinstance(self.evidence, list):
for i, evid in enumerate(self.evidence):
mcASPObj = marginal_mcsat.mcSAT(finalout, evid, [],
self.xorMode,
self.max_mcsat_iteration)
mcASPObj.args = self.args
mcASPObj.runMCASP()
samplesEvidenced = mcASPObj.sampleForReturn
if self.args.verbosity > 4:
print("Done sampling from " + str(i) +
"th partial evidence")
for sampleE in samplesEvidenced:
for atom in sampleE:
if (str(atom.name) == "neg" or str(atom.name)
== "unsat") and "lpmlnneg_" in str(
atom.arguments[0]):
idx = eval(str(atom.arguments[0]).split('_')[1])
for key, value in self.weightsDic.items():
if value['wIndex'] == idx:
value['nii'] += 1
for key, value in self.weightsDic.items():
value['nii'] = float(value['nii']) / float(
self.max_mcsat_iteration)
else:
mcASPObj = marginal_mcsat.mcSAT(finalout, self.evidence, [],
self.xorMode,
self.max_mcsat_iteration)
mcASPObj.args = self.args
mcASPObj.runMCASP()
samplesEvidenced = mcASPObj.sampleForReturn
if self.args.verbosity > 4:
print("Done samples evidenced")
for sampleE in samplesEvidenced:
for atom in sampleE:
if (str(atom.name) == "neg" or str(atom.name)
== "unsat") and "lpmlnneg_" in str(
atom.arguments[0]):
idx = eval(str(atom.arguments[0]).split('_')[1])
for key, value in self.weightsDic.items():
if value['wIndex'] == idx:
value['nii'] += 1
for key, value in self.weightsDic.items():
value['nii'] = float(value['nii']) / float(
self.max_mcsat_iteration)
def max_walk_sat_exact_sdd(self,save=None,load=None):
solution = ""
utility = -float("inf")
inferObj = sdd_infer.sddInference("", ['utility'])
inferObj.args = self.args
if load is None and save is None:
inferObj.sddConstructorFromLPMLN(self.aspContent)
elif load is None and save is not None:
inferObj.sddConstructorFromLPMLN(self.aspContent,True,save)
else:
inferObj.sddConstructorFromFile(load)
for _ in range(1,self.mt):
mcASPObj = marginal_mcsat.mcSAT(self.aspContent, "", [], self.xorMode,1)
mcASPObj.runMCASP()
firstSample = mcASPObj.getSamples()
decisionAssignments = self.completeTruthAssignment(firstSample[0])
tempSoln = decisionAssignments
tempUtility = self.max_walk_sat_exact_helper(inferObj,tempSoln)
for _ in range(1,self.mf):
vf=""
deltCostVf = float("inf")
orgUtility = self.max_walk_sat_exact_helper(inferObj,tempSoln)
if self.p < random.random():
vf = tempSoln[random.randint(0, len(tempSoln) - 1)]
newDecisionAssignments = self.flipDecisionAtom(tempSoln,vf)
flippedUtility =self.max_walk_sat_exact_helper(inferObj,newDecisionAssignments)
if flippedUtility != False:
deltCostVf = orgUtility - flippedUtility
else:
deltaCost = []
for decToFlip in tempSoln:
newDecisionAssignments = self.flipDecisionAtom(tempSoln, decToFlip)
flippedUtility = self.max_walk_sat_exact_helper(inferObj,newDecisionAssignments)
if flippedUtility != False:
deltaCost.append((decToFlip,orgUtility - flippedUtility))
if len(deltaCost) != 0:
minDeltaCost = min(deltaCost,key=lambda t:t[1])
else:
continue
vf = minDeltaCost[0]
deltCostVf = minDeltaCost[1]
if deltCostVf<0:
tempSoln = self.flipDecisionAtom(tempSoln, vf)
tempUtility = tempUtility - deltCostVf
if tempUtility >utility:
utility = tempUtility
solution = tempSoln
return solution,utility
def max_walk_sat(self):
solution = ""
utility = -float("inf")
for mt_iter in range(1,self.mt):
if self.args.verbosity >4:
print("Starting one iteration: ",mt_iter)
mcASPObj = marginal_mcsat.mcSAT(self.aspContent, "", [], self.xorMode,1)
mcASPObj.runMCASP()
firstSample = mcASPObj.getSamples()
decisionAssignments = self.completeTruthAssignment(firstSample[0])
tempSoln = decisionAssignments
tempUtility = self.expectedUtility(self.buildASPEvidenceForm(tempSoln))
for _ in range(1,self.mf):
vf=""
deltCostVf = float("inf")
orgUtility = self.expectedUtility(self.buildASPEvidenceForm(tempSoln))
if self.p < random.random():
vf = tempSoln[random.randint(0, len(tempSoln) - 1)]
newDecisionAssignments = self.flipDecisionAtom(tempSoln,vf)
flippedUtility = self.expectedUtility(self.buildASPEvidenceForm(newDecisionAssignments))
if flippedUtility != False:
deltCostVf = orgUtility - flippedUtility
else:
deltaCost = []
for decToFlip in tempSoln:
newDecisionAssignments = self.flipDecisionAtom(tempSoln, decToFlip)
flippedUtility = self.expectedUtility(self.buildASPEvidenceForm(newDecisionAssignments))
if flippedUtility != False:
deltaCost.append((decToFlip,orgUtility - flippedUtility))
if len(deltaCost) != 0:
minDeltaCost = min(deltaCost,key=lambda t:t[1])
else:
continue
vf = minDeltaCost[0]
deltCostVf = minDeltaCost[1]
if deltCostVf<0:
tempSoln = self.flipDecisionAtom(tempSoln, vf)
tempUtility = tempUtility - deltCostVf
if tempUtility >utility:
utility = tempUtility
solution = tempSoln
return solution,utility
def max_walk_sat_exact_sdd_2(self):
solution = ""
utility = -float("inf")
r = run.run(self.aspContent,self.args)
for _ in range(1,self.mt):
mcASPObj = marginal_mcsat.mcSAT(self.aspContent, "", [], self.xorMode,1)
mcASPObj.runMCASP()
firstSample = mcASPObj.getSamples()
decisionAssignments = self.completeTruthAssignment(firstSample[0])
tempSoln = decisionAssignments
tempUtility = self.max_walk_sat_sdd2_helper(r, tempSoln)
for _ in range(1,self.mf):
vf=""
deltCostVf = float("inf")
orgUtility = self.max_walk_sat_sdd2_helper(r,tempSoln)
if self.p < random.random():
vf = tempSoln[random.randint(0, len(tempSoln) - 1)]
newDecisionAssignments = self.flipDecisionAtom(tempSoln,vf)
flippedUtility =self.max_walk_sat_sdd2_helper(r,newDecisionAssignments)
if flippedUtility != False:
deltCostVf = orgUtility - flippedUtility
else:
deltaCost = []
for decToFlip in tempSoln:
newDecisionAssignments = self.flipDecisionAtom(tempSoln, decToFlip)
flippedUtility = self.max_walk_sat_sdd2_helper(r,newDecisionAssignments)
if flippedUtility != False:
deltaCost.append((decToFlip,orgUtility - flippedUtility))
if len(deltaCost) != 0:
minDeltaCost = min(deltaCost,key=lambda t:t[1])
else:
continue
vf = minDeltaCost[0]
deltCostVf = minDeltaCost[1]
if deltCostVf<0:
tempSoln = self.flipDecisionAtom(tempSoln, vf)
tempUtility = tempUtility - deltCostVf
if tempUtility >utility:
utility = tempUtility
solution = tempSoln
return solution,utility
def expectedUtility(self,query_action):
mcASPObj = marginal_mcsat.mcSAT(self.aspContent, query_action, [], self.xorMode,self.max_sample)
mcASPObj.args = self.args
mcASPObj.runMCASP()
samplesEvidenced = mcASPObj.getSamples()
#Given program combining with decision or evidence cannot lead to any stable model
if len(samplesEvidenced) == 0:
return False
expectedU = 0.0
for sampleE in samplesEvidenced:
for atom in sampleE:
if str(atom.name) == "utility":
expectedU+=eval(str(atom.arguments[0]).strip('\"'))
return expectedU/len(samplesEvidenced)
def findTotalMisWithMCSAT(self, finalout):
for key, value in self.weightsDic.items():
value['unsatEvi'] = 0
value['unsatNoEvi'] = 0
if isinstance(self.evidence, list):
for i, evid in enumerate(self.evidence):
if self.args.verbosity > 4:
print("Start sample no evidenced")
mcASPObj = marginal_mcsat.mcSAT(finalout, "", [], self.xorMode,
self.max_mcsat_iteration)
mcASPObj.args = self.args
mcASPObj.runMCASP()
samples = mcASPObj.sampleForReturn
if self.args.verbosity > 4:
print("Done sample no evidenced")
for samplene in samples:
for atom in samplene:
counter = 0
if (atom.name == "neg" or atom.name == "unsat") and (
"lpmlnneg_" in str(atom.arguments[0])):
idx = eval(str(atom.arguments[0]).split('_')[1])
counter += 1
for key, value in self.weightsDic.items():
if value['wIndex'] == idx:
value['unsatNoEvi'] += 1
if self.args.verbosity > 4:
if self.complete_evid:
print("Start sampling from " + str(i) +
"th complete evidenced")
else:
print("Start sampling from " + str(i) +
"th partial evidenced")
if self.complete_evid:
mcASPObj = marginal_mcsat.mcSAT(finalout, evid, [],
self.xorMode, 1)
else:
mcASPObj = marginal_mcsat.mcSAT(finalout, evid, [],
self.xorMode,
self.max_mcsat_iteration)
mcASPObj.args = self.args
mcASPObj.runMCASP()
samplesEvidenced = mcASPObj.sampleForReturn
if self.args.verbosity > 4:
if self.complete_evid:
print("Done sampling from " + str(i) +
"th complete evidenced")
else:
print("Done sampling from " + str(i) +
"th partial evidenced")
for sampleE in samplesEvidenced:
for atom in sampleE:
if (str(atom.name) == "neg" or str(atom.name)
== "unsat") and "lpmlnneg_" in str(
atom.arguments[0]):
idx = eval(str(atom.arguments[0]).split('_')[1])
for key, value in self.weightsDic.items():
if value['wIndex'] == idx:
if self.complete_evid:
value[
'unsatEvi'] += 1 * self.max_mcsat_iteration
else:
value['unsatEvi'] += 1
else:
if self.args.verbosity > 4:
if self.complete_evid:
print("Start sampling from complete evidenced")
else:
print("Start sampling from partial evidenced")
if self.complete_evid:
mcASPObj = marginal_mcsat.mcSAT(finalout, self.evidence, [],
self.xorMode, 1)
else:
mcASPObj = marginal_mcsat.mcSAT(finalout, self.evidence, [],
self.xorMode,
self.max_mcsat_iteration)
mcASPObj.args = self.args
mcASPObj.runMCASP()
samplesEvidenced = mcASPObj.sampleForReturn
if self.args.verbosity > 4:
if self.complete_evid:
print("Done sample from complete evidenced")
else:
print("Done sample from partial evidenced")
for sampleE in samplesEvidenced:
for atom in sampleE:
if (str(atom.name) == "neg" or str(atom.name)
== "unsat") and "lpmlnneg_" in str(
atom.arguments[0]):
idx = eval(str(atom.arguments[0]).split('_')[1])
for key, value in self.weightsDic.items():
if value['wIndex'] == idx:
if self.complete_evid:
value[
'unsatEvi'] += 1 * self.max_mcsat_iteration
else:
value['unsatEvi'] += 1
if self.args.verbosity > 4:
print(self.weightsDic)
def learn(self):
self.learn_ini()
content = self.lpmln_to_lpmln_neg_parser(
self.progranWithoutPlaceholder)
content = self.lpmln_to_asp_parser(content)
finalout = self.asp_domain_2_asp_parser(content)
if isinstance(self.evidence, list):
for l in self.evidence:
mcASPObj = marginal_mcsat.mcSAT(finalout, l, [], self.xorMode,
self.min_mcsat_interation)
mcASPObj.args = self.args
result = mcASPObj.runMCASP()
if not result:
print('Evidence and program not satisfiable. Exit.')
return False
else:
mcASPObj = marginal_mcsat.mcSAT(finalout, l, [], self.xorMode,
self.min_mcsat_interation)
mcASPObj.args = self.args
result = mcASPObj.runMCASP()
if not result:
print('Evidence and program not satisfiable. Exit.')
return False
# Begin: Learning Iterations
actualNumIteration = 0
for iter_count in range(self.max_learning_iteration):
iter_start_time = time.time()
actualNumIteration += 1
self.updateWithWeightPlaceHolders()
content = self.lpmln_to_lpmln_neg_parser(
self.progranWithoutPlaceholder)
content = self.lpmln_to_asp_parser(content)
finalout = self.asp_domain_2_asp_parser(content)
self.findTotalMisWithMCSAT(finalout)
# End: Single learning iteration
# Compute new weights
total_gradient = 0
max_diff = 0
for key, value in self.weightsDic.items():
prob_gradient = float(-value['unsatEvi'] +
float(value['unsatNoEvi'])) / float(
self.max_mcsat_iteration)
total_gradient += abs(prob_gradient)
if self.args.verbosity > 4:
print('Rule: ', key)
print(
'# False ground instances from Evidence: ',
float(value['unsatEvi']) /
float(self.max_mcsat_iteration))
print(
'Expected # false ground instances: ',
float(value['unsatNoEvi']) /
float(self.max_mcsat_iteration))
print('Gradient: ', prob_gradient)
if max_diff < abs(self.lr * prob_gradient):
max_diff = abs(self.lr * prob_gradient)
value['weight'] += (self.lr * prob_gradient)
value['weightSum'] += value['weight']
if self.args.verbosity > 4:
print('New weight: ', key, ':', value['weight'])
# End: Learning Iterations
if self.args.verbosity > 4:
print("max_diff: ", max_diff)
if max_diff <= self.stopping_diff:
break
if self.args.verbosity > 4:
print('============ Time for finishing iteration ' +
str(actualNumIteration) + ': ' +
str(time.time() - iter_start_time) + ' ============')
# Begin: Store and save new weights
self.updateWithWeightPlaceHolders()
self.updatefinalWeight()
def solve(self):
wholeInput = ""
evidenceInput = ""
if self.arg_list['nn_indictor']:
wholeInput += nn_processor.process(self.arg_list['file_name'][-1])
self.arg_list['file_name'] = self.arg_list['file_name'][:-1]
for lpmln_file in self.arg_list['file_name']:
with open(lpmln_file, 'r') as lpmln_content:
wholeInput += lpmln_content.read()
lpmln_content.close()
for evid_file in self.arg_list['evidence']:
with open(evid_file, 'r') as evidence_content:
evidenceInput += evidence_content.read()
evidence_content.close()
if self.arg_list['mode'] == "sdd":
content = self.lpmln_to_asp_sdd_parser(wholeInput, False,
self.arg_list['mf'])
parsed_lpmln = self.asp_domain_2_asp_parser(content)
elif self.arg_list['mode'] == "mcasp":
content = self.lpmln_to_lpmln_neg_parser(wholeInput,
self.arg_list['mf'])
content = self.lpmln_to_asp_parser(content, False,
self.arg_list['mf'])
parsed_lpmln = self.asp_domain_2_asp_parser(content)
elif self.arg_list['mode'] == "map":
content = self.lpmln_to_asp_parser(wholeInput,
self.arg_list['hard'],
self.arg_list['mf'], True)
parsed_lpmln = self.asp_domain_2_asp_parser(content)
else:
content = self.lpmln_to_asp_parser(wholeInput,
self.arg_list['hard'],
self.arg_list['mf'], True)
parsed_lpmln = self.asp_domain_2_asp_parser(content)
if self.args.verbosity > 4:
print(
"================== Parsed ASP Program ======================")
print(parsed_lpmln)
if self.arg_list['mode'] == "map":
with open('asp.pl', 'w') as fw:
fw.write(parsed_lpmln + evidenceInput)
if self.arg_list['hard']:
command = "clingo " + os.getcwd() + "/asp.pl --opt-mode=enum 0"
else:
command = "clingo " + os.getcwd() + "/asp.pl "
os.system(command)
elif self.arg_list['mode'] == "ex":
warn_option = "--warn=none"
thread_option = "-t 4"
enumerateAll = "--opt-mode=enum"
listAll = "0"
warn = "--warn=no-atom-undefined"
options = [warn_option, thread_option, enumerateAll, listAll, warn]
exactInfer = exactProbInfer.exactProbInfer(
parsed_lpmln + evidenceInput, options, self.arg_list['query'],
self.arg_list['hard'], self.arg_list['all'])
if self.arg_list['hard']:
with open('asp.pl', 'w') as fw:
fw.write(parsed_lpmln + evidenceInput)
command = "clingo " + os.getcwd() + "/asp.pl --opt-mode=enum 0"
os.system(command)
exactInfer.solve()
elif self.arg_list['mode'] == "mcasp":
mcASPObj = marginal_mcsat.mcSAT(parsed_lpmln, evidenceInput,
self.arg_list['query'],
self.arg_list['xor_mode'],
self.arg_list['sample'])
mcASPObj.args = self.args
mcASPObj.runMCASP()
mcASPObj.printQuery()
elif self.arg_list['mode'] == "kcor":
marginal_coherent_obj = marginal_coherent.Marginal_coherent(
parsed_lpmln, evidenceInput, self.arg_list['query'],
self.arg_list['sample'])
marginal_coherent_obj.args = self.args
marginal_coherent_obj.run()
marginal_coherent_obj.printQuery()
elif self.arg_list['mode'] == "sdd":
inferObj = sdd_infer.sddInference(evidenceInput,
self.arg_list['query'])
inferObj.args = self.args
if self.arg_list['sddPara'][0] == 0:
inferObj.sddConstructorFromLPMLN(parsed_lpmln, False)
inferObj.inferencePrtQuery()
elif self.arg_list['sddPara'][0] == 1:
inferObj.sddConstructorFromFile(self.arg_list['sddPara'][1])
inferObj.inferencePrtQuery()
else:
inferObj.sddConstructorFromLPMLN(parsed_lpmln, True,
self.arg_list['sddPara'][1])
inferObj.inferencePrtQuery()
else:
r = run.run(parsed_lpmln, self.args)
result = r.infer(self.arg_list['query'], evidenceInput)
for r in result:
print(r[0].to_string(), ": ", r[1])
