def _computeCounts(self):
''' compute soft counts (assuming independence) '''
allSoft = self.params.get("allSoft", False)
if allSoft == False:
# compute regular counts for all "normal" possible worlds
LL._computeCounts(self)
# add another world for soft beliefs
baseWorld = self.mrf.worlds[self.idxTrainingDB]
self.mrf.worlds.append({"values": baseWorld["values"]})
self.idxTrainingDB = len(self.mrf.worlds) - 1
# and compute soft counts only for that world
softCountWorldIndices = [self.idxTrainingDB]
else:
# compute soft counts for all possible worlds
self.counts = {}
softCountWorldIndices = xrange(len(self.mrf.worlds))
# compute soft counts
for i in softCountWorldIndices:
world = self.mrf.worlds[i]
if i == self.idxTrainingDB:
print "TrainingDB: prod, groundformula"
for gf in self.mrf.gndFormulas:
prod = truthDegreeGivenSoftEvidence(gf, world["values"], self.mrf)
key = (i, gf.idxFormula)
cnt = self.counts.get(key, 0)
cnt += prod
self.counts[key] = cnt
if i == self.idxTrainingDB:
print "%f gf: %s" % (prod, str(gf))
print "worlds len: ", len(self.mrf.worlds)
def _computeCounts(self):
''' compute soft counts (assuming independence) '''
allSoft = self.params.get("allSoft", False)
if allSoft == False:
# compute regular counts for all "normal" possible worlds
LL._computeCounts(self)
# add another world for soft beliefs
baseWorld = self.mrf.worlds[self.idxTrainingDB]
self.mrf.worlds.append({"values": baseWorld["values"]})
self.idxTrainingDB = len(self.mrf.worlds) - 1
# and compute soft counts only for that world
softCountWorldIndices = [self.idxTrainingDB]
else:
# compute soft counts for all possible worlds
self.counts = {}
softCountWorldIndices = xrange(len(self.mrf.worlds))
# compute soft counts
for i in softCountWorldIndices:
world = self.mrf.worlds[i]
if i == self.idxTrainingDB:
print "TrainingDB: prod, groundformula"
for gf in self.mrf.gndFormulas:
prod = truthDegreeGivenSoftEvidence(gf, world["values"],
self.mrf)
key = (i, gf.idxFormula)
cnt = self.counts.get(key, 0)
cnt += prod
self.counts[key] = cnt
if i == self.idxTrainingDB:
print "%f gf: %s" % (prod, str(gf))
print "worlds len: ", len(self.mrf.worlds)
def _getTruthDegreeGivenEvidence(self, gf, worldValues=None):
if worldValues is None: worldValues = self.mrf.evidence
return truthDegreeGivenSoftEvidence(gf, worldValues, self.mrf)
def _getTruthDegreeGivenEvidence(self, gf, worldValues=None):
if worldValues is None: worldValues = self.mrf.evidence
return truthDegreeGivenSoftEvidence(gf, worldValues, self.mrf)
def _f(self, wt):
self._calculateWorldValues(wt) #only to calculate partition function here:
#self._calculateWorldProbabilities()
#new idea: minimize squared error of world prob. given by weights and world prob given by soft evidence
error = 0
#old method (does not work with mixed hard and soft evidence)
if True:
for idxWorld, world in enumerate(self.mrf.worlds):
if idxWorld in self.worldProbabilities: #lambda_x
worldProbability = self.worldProbabilities[idxWorld]
else:
worldProbability = 0
worldProbGivenWeights = self.expsums[idxWorld] / self.partition_function
error += abs(worldProbGivenWeights - worldProbability)
#print "worldProbGivenWeights - worldProbability ", worldProbGivenWeights, "-", worldProbability
# for idxWorld, worldProbability in self.worldProbabilities.iteritems(): #lambda_x
# worldProbGivenWeights = self.expsums[idxWorld] / self.partition_function
# error += abs(worldProbGivenWeights - worldProbability)
# #print "world:", self.mrf.worlds[idxWorld]
# print "worldProbGivenWeights - worldProbability ", worldProbGivenWeights, "-", worldProbability
if False:#new try, doesn't work...
for idxWorld, world in enumerate(self.mrf.worlds):
worldProbGivenWeights = self.expsums[idxWorld] / self.partition_function
#compute countDiffSum:
#for i, world in enumerate(self.mrf.worlds):
if idxWorld not in self.countsByWorld:
print "computing counts for:", idxWorld
counts = {} #n
for gf in self.mrf.gndFormulas:
if self.mrf._isTrue(gf, self.mrf.worlds[idxWorld]["values"]):
key = gf.idxFormula
cnt = counts.get(key, 0)
cnt += 1
counts[key] = cnt
self.countsByWorld[idxWorld] = counts
#� (soft counts for evidence)
if len(self.softCountsEvidenceWorld) == 0:
print "computing evidence soft counts"
self.softCountsEvidenceWorld = {}
for gf in self.mrf.gndFormulas:
prod = truthDegreeGivenSoftEvidence(gf, self.mrf.evidence, self.mrf)
key = gf.idxFormula
cnt = self.softCountsEvidenceWorld.get(key, 0)
cnt += prod
self.softCountsEvidenceWorld[key] = cnt
#if i == self.idxTrainingDB:
# print "%f gf: %s" % (prod, str(gf))
countDiffSum = 0
for idxFormula, count in self.countsByWorld[idxWorld].iteritems():
countDiffSum += abs(count - self.softCountsEvidenceWorld[idxFormula])
#print "countDiffSum", countDiffSum, "worldProbability", worldProbGivenWeights
error += worldProbGivenWeights * ((countDiffSum)**2)
print "wt =", wt
print "error:", error
ll = -error
print
return ll
def _f(self, wt):
self._calculateWorldValues(
wt) #only to calculate partition function here:
#self._calculateWorldProbabilities()
#new idea: minimize squared error of world prob. given by weights and world prob given by soft evidence
error = 0
#old method (does not work with mixed hard and soft evidence)
if True:
for idxWorld, world in enumerate(self.mrf.worlds):
if idxWorld in self.worldProbabilities: #lambda_x
worldProbability = self.worldProbabilities[idxWorld]
else:
worldProbability = 0
worldProbGivenWeights = self.expsums[
idxWorld] / self.partition_function
error += abs(worldProbGivenWeights - worldProbability)
#print "worldProbGivenWeights - worldProbability ", worldProbGivenWeights, "-", worldProbability
# for idxWorld, worldProbability in self.worldProbabilities.iteritems(): #lambda_x
# worldProbGivenWeights = self.expsums[idxWorld] / self.partition_function
# error += abs(worldProbGivenWeights - worldProbability)
# #print "world:", self.mrf.worlds[idxWorld]
# print "worldProbGivenWeights - worldProbability ", worldProbGivenWeights, "-", worldProbability
if False: #new try, doesn't work...
for idxWorld, world in enumerate(self.mrf.worlds):
worldProbGivenWeights = self.expsums[
idxWorld] / self.partition_function
#compute countDiffSum:
#for i, world in enumerate(self.mrf.worlds):
if idxWorld not in self.countsByWorld:
print "computing counts for:", idxWorld
counts = {} #n
for gf in self.mrf.gndFormulas:
if self.mrf._isTrue(
gf, self.mrf.worlds[idxWorld]["values"]):
key = gf.idxFormula
cnt = counts.get(key, 0)
cnt += 1
counts[key] = cnt
self.countsByWorld[idxWorld] = counts
#� (soft counts for evidence)
if len(self.softCountsEvidenceWorld) == 0:
print "computing evidence soft counts"
self.softCountsEvidenceWorld = {}
for gf in self.mrf.gndFormulas:
prod = truthDegreeGivenSoftEvidence(
gf, self.mrf.evidence, self.mrf)
key = gf.idxFormula
cnt = self.softCountsEvidenceWorld.get(key, 0)
cnt += prod
self.softCountsEvidenceWorld[key] = cnt
#if i == self.idxTrainingDB:
# print "%f gf: %s" % (prod, str(gf))
countDiffSum = 0
for idxFormula, count in self.countsByWorld[
idxWorld].iteritems():
countDiffSum += abs(
count - self.softCountsEvidenceWorld[idxFormula])
#print "countDiffSum", countDiffSum, "worldProbability", worldProbGivenWeights
error += worldProbGivenWeights * ((countDiffSum)**2)
print "wt =", wt
print "error:", error
ll = -error
print
return ll
