def CreateBuckets(self, trainSet, testSet, vSets, name, testCut, iternum,
save, superbit, stages, buckets, doprint, clfName,
tunelrn):
out = []
if self.isCount:
keySet = list(
DPLIB.getMeasuresCount([0, 1, 2, 3], [0, 1, 2, 3]).keys())
else:
keySet = list(
DPLIB.getExtMeasures({
"tp": 1,
"tn": 2,
"fp": 3,
"fn": 4
}).keys())
out.append("#STARTED FOR-" + name + ":" + self.file + ": ")
startTime = Common.getCurrentTimeMil()
spentIsTime = 0
tempTime = 0
out.append("#Using also Label For train in LSH")
if (vSets == None):
vSets = []
vSets.append(trainSet)
if (save):
DPLIB.SaveToCsv(
trainSet, "MAIN-TRAIN-FILE-" + "ITER=" + str(iternum) + "--" +
"METHOD=" + name + "--FILE=" + self.file + "--")
DPLIB.SaveToCsv(
testSet, "MAIN-TEST-FILE-" + "ITER=" + str(iternum) + "--" +
"METHOD=" + name + "--FILE=" + self.file + "--")
for i in range(len(vSets)):
DPLIB.SaveToCsv(
trainSet,
"VSET-FILE-" + "INDEX=" + str(i) + "ITER=" + str(iternum) +
"--" + "METHOD=" + name + "--FILE=" + self.file + "--")
np.random.shuffle(trainSet)
np.random.shuffle(testSet)
tempTime = Common.getCurrentTimeMil()
count = len(trainSet)
bins = {}
# R^n
n = trainSet.shape[1] - 1
binid = 0
#lshmin = LSHMinHash(stages, buckets, n);
try:
lshsuper = LSHSuperBit(stages=stages,
buckets=buckets,
dimensions=n)
except Exception as ex:
print('##SuperBit with specified parameters failed:' + str(ex))
return None
sp = 0.75
# Compute a SuperBit signature, and a LSH hash
for i in range(count):
vector = trainSet[i, 1:].tolist()
hash = None
if (superbit):
hash = lshsuper.hash(vector)
else:
##Minhash support
# #hash = lshmin.hash(vecBool);
pass
binid = hash[0]
if not binid in bins.keys():
bins[binid] = []
bins[binid].append(trainSet[i])
spentIsTime += Common.getCurrentTimeMil() - tempTime
numBins = len(bins.keys())
for binid in bins.keys():
bins[binid] = np.array(bins[binid])
out.append("#Number of BINS:" + name + ":" + self.file + ": " +
str(numBins))
pop = []
for i in bins.keys():
trSet = bins[i]
l = GLOB(clfName, tunelrn).getClassifier()
#if (tunelrn):
# l = l.getTunedCLF(trSet, vSets,fout,name, file);
l.buildClassifier(trSet)
cf = 0
j = 0
allvecs = []
confs = []
allcfs = []
allaucs = []
valsA = None
confsA = None
aucA = 0.0
for vSet in vSets:
vec = None
actuals = None
vec = l.evaluateModel(vSet)
actuals = vSet[:, -1]
vals = None
auc = 0
if self.isCount:
vals = DPLIB.getMeasuresCount(actuals, vec)
else:
auc = DPLIB.getAUC(actuals, vec)
aucA += auc
allaucs.append(auc)
if (testCut):
vCF = 0.1
bestCF = 0
bestCFVal = -1
bestVals = None
while True:
tvals = DPLIB.getConfMatrix(actuals, vec, vCF)
measures = DPLIB.getMeasures(tvals)
fit = measures["F"] * measures["GMean1"]
if (fit > bestCFVal or bestVals == None):
bestCFVal = fit
bestCF = vCF
bestVals = tvals
vCF += 0.1
if (vCF >= 1):
break
if (confsA == None):
confsA = {key: 0 for key in bestVals.keys()}
for j in confsA.keys():
confsA[j] += bestVals[j]
confs.append(bestVals)
vals = DPLIB.getMeasures(bestVals)
cf += bestCF
allcfs.append(bestCF)
else:
tvals = DPLIB.getConfMatrix(actuals, vec)
if (confsA == None):
confsA = {key: 0 for key in tvals.keys()}
for j in confsA.keys():
confsA[j] += tvals[j]
confs.append(tvals)
vals = DPLIB.getMeasures(tvals)
allcfs.append(DPLIB.DefaultCF)
allvecs.append(vals)
if (valsA == None):
valsA = {key: 0 for key in keySet}
for j in keySet:
valsA[j] += vals[j]
for j in keySet:
valsA[j] /= len(vSets)
h = None
if not self.isCount:
for j in confsA.keys():
confsA[j] /= len(vSets)
if (testCut):
cf /= len(vSets)
aucA /= len(vSets)
h = CHRM_GIS(trSet, valsA, aucA)
h.fitnesses = allvecs
h.aucs = allaucs
h.conf = confsA
h.confs = confs
h.allcfs = allcfs
if (testCut):
h.bestCF = cf
else:
h.bestCF = DPLIB.DefaultCF
else:
h = CHRM_GIS_Count(trSet, valsA)
h.fitnesses = allvecs
pop.append(h)
l = None
tempTime = Common.getCurrentTimeMil()
pop = DPLIB.MySort(pop)
spentIsTime += Common.getCurrentTimeMil() - tempTime
top = pop[0]
out.append("#Instances in Top:" + str(len(top.ds)))
out.append("#STAGES:" + name + ":" + self.file + ": " + str(stages))
out.append("#BUCKETS:" + name + ":" + self.file + ": " + str(buckets))
if not self.isCount:
out.append("#BEST-CF-VALUE:" + name + ":" + self.file + ": " +
str(top.bestCF))
l = GLOB(clfName, tunelrn).getClassifier()
if (tunelrn):
l = l.getTunedCLF(top.ds, vSets, fout, name, file)
out.append("#TUNE-LRN-PARAMS-" + name + ":" + self.file + ": " +
str(l.selectedParams))
sCheck = l.getCLFOptions()
out.append("#SETSET-LRN-PARAMS-" + name + ":" + self.file + ": " +
str(sCheck))
l.buildClassifier(top.ds)
vec = l.evaluateModel(testSet)
out.append("#LSH-FOR-TOP-ONLY")
if self.isCount:
vals = DPLIB.getMeasuresCount(testSet[:, -1], vec)
out.append(name + ":" + self.file + ": " + str(vals))
else:
tvals = DPLIB.getConfMatrix(testSet[:, -1], vec, top.bestCF)
out.append("#CONF-TEST-" + name + ":" + self.file + ": " +
str(tvals))
vals = DPLIB.getMeasures(tvals)
auc = DPLIB.getAUC(testSet[:, -1], vec)
vals['auc'] = auc
out.append(name + ":" + self.file + ": " + str(vals))
for i in range(len(pop)):
pop[i] = None
pop = None
for i in bins.keys():
bins[i] = None
bins = None
time = Common.getCurrentTimeMil() - startTime
if (name.find("LSHTune") < 0):
out.append("#TIME-FOR:" + name + ":" + self.file + ": " +
str(time))
out.append("#TIME-FOR-IS:" + name + ":" + self.file + ": " +
str(spentIsTime))
self.output = +out
top.addToExtra("SPENT-TIME-IS", float(spentIsTime))
return top, out
def run(self, trainSeti, testSeti, name, fout, vSets, vSetType,
fixedTrainSize, log, ignoreOK, threshold, thresholds, rejectedFits,
rejectedPerfs, rejectedTestPerfs, clfName):
mad = 0.0
if self.isCount:
keySet = list(
DPLIB.getMeasuresCount([0, 1, 2, 3], [0, 1, 2, 3]).keys())
mad = DPLIB.SetBugCountForMut(trainSeti)
else:
keySet = list(
DPLIB.getExtMeasures({
"tp": 1,
"tn": 2,
"fp": 3,
"fn": 4
}).keys())
startTime = Common.getCurrentTimeMil()
tempTime = 0
spentISTime = 0
#For Binary Prediction, isCount = False
auc = 0
preds = []
pop = []
trainSet = np.copy(trainSeti)
testSet = np.copy(testSeti)
pop.clear()
tstSize = len(testSet)
partSize = int(tstSize / self.numParts)
preds.clear()
diffs = []
auc = 0.0
#For isCount = True
actuals = []
prrs = []
if (log):
self.prnt("#GIS-OPTIONS;;For=" + name + "@" + ":iters=" +
str(self.iters) + "-POPSIZE=" + str(self.popSize) +
"-NumParts=" + str(self.numParts) + "-NumGens=" +
str(self.numGens) + "-sizeTop=" + str(self.sizeTopP) +
"-Learner=" + clfName + "\n")
isOK = True
np.random.shuffle(testSet)
self.FinalLearners = []
self.FinalDatasets = []
for p in range(self.numParts):
diffp = []
self.prnt("\n" + str(p) + ": ")
tempTime = Common.getCurrentTimeMil()
pop.clear()
start = p * partSize
end = (p + 1) * partSize
if (end > tstSize):
end = tstSize
if (p == self.numParts - 1):
end = tstSize
testPart = testSet[start:end, :]
spentISTime += Common.getCurrentTimeMil() - tempTime
uinds = set()
if (vSets == None or len(vSets) == 0):
if (vSets == None):
vSets = []
vSet = None
retVal = ""
if (vSetType == 'Train Set'):
vSet = trainSeti
if (log):
retVal = DPLIB.getStats(vSet, True, True, True)
self.prnt("#VSETINFO;;prt=" + str(p) + ";;For=" +
name + "@" + ":" + retVal + "\n")
retVal = None
elif (vSetType == 'NN-Filter'):
tempTime = Common.getCurrentTimeMil()
vSet = DPLIB.NNFilter(trainSet, testPart, 1)
spentISTime += Common.getCurrentTimeMil() - tempTime
if (log):
retVal = DPLIB.getStats(vSet, True, True, True)
self.prnt("#VSETINFO;;prt=" + str(p) + ";;For=" +
name + "@" + ":" + retVal + "\n")
retVal = None
#If random, but not fed into the func, generate one randomly, with size of testPart
elif (vSetType == 'Multiple Random'
or vSetType == 'Single Random'):
size = len(testPart)
vSet = []
j = 0
while (j < size):
index = np.random.randint(trainSet.numInstances())
if (not index in uinds):
uinds.add(index)
else:
continue
vSets.append(trainSet[index])
j += 1
if (log):
retVal = DPLIB.getStats(vSet, true, true, True)
self.prnt("#VSETINFO;;prt=" + str(p) + ";;For=" +
name + "@" + ":" + retVal + "\n")
retVal = None
vSet = np.array(vSet)
elif (vSetType == '!!TEST!!'):
#Upper Bound. Should not be used.
self.prnt("Should not be used.")
vSet = testSeti
if (log):
retVal = DPLIB.getStats(vSet, True, True, True)
self.prnt("#VSETINFO;;prt=" + str(p) + ";;For=" +
name + "@" + ":" + retVal + "\n")
retVal = None
elif vSetType == 'KS2':
vSet = None
vSets.append(vSet)
else:
retVal = ""
for vSet in vSets:
if (log):
retVal = DPLIB.getStats(vSet, True, True, True)
self.prnt("#VSETINFO;;prt=" + str(p) + ";;For=" +
name + "@" + ":" + retVal + "\n")
retVal = None
for i in range(self.popSize):
tempTime = Common.getCurrentTimeMil()
uinds.clear()
size = 0
if (fixedTrainSize):
size = self.chrmSize
else:
size = np.random.randint(self.chrmSize) + 10
while True:
trSet = []
j = 0
while (j < size):
index = np.random.randint(len(trainSet))
trSet.append(trainSet[index])
if (not index in uinds):
uinds.add(index)
j += 1
spentISTime += Common.getCurrentTimeMil() - tempTime
trSet = np.array(trSet)
if len(set(list(trSet[:, -1]))) >= 2:
break
tempTime = Common.getCurrentTimeMil()
pv, p_vals = DPLIB.checkSimilarity(trSet[:, :-1],
testPart[:, :-1])
if self.isCount:
h = CHRM_GIS_Count(trSet, None, extraAsFitness='p-val')
h.addToExtra('p-val', sum(p_vals))
pop.append(h)
else:
h = CHRM_GIS(trSet, None, None, extraAsFitness='p-val')
h.addToExtra('p-val', sum(p_vals))
pop.append(h)
spentISTime += Common.getCurrentTimeMil() - tempTime
tempTime = Common.getCurrentTimeMil()
pop = DPLIB.MySort(pop)
spentISTime += Common.getCurrentTimeMil() - tempTime
cnt = 0
g = 0
for g in range(self.numGens):
self.prnt(str(g) + " ")
if (log):
pass
#retVal = ""
#for i in range(len(pop)):
# chrm = pop[i]
# retVal = DPLIB.getStats(chrm.ds, False, False, False);
# self.prnt("#POPITNFO;;gn="+str(g)+";;prt="+str(p)+";;For="+name+"@"+":"+retVal+"\n");
# self.prnt("#POPITVALS;;gn="+str(g)+";;prt="+str(p)+";;For="+name+"@"+":"+"rpaf="+str(chrm.fitness).replace(", ", ",")
# +";;conf="+str(chrm.conf).replace(", ", ",")+";;fit="+str(chrm.getFitness())+";;TConf2="+str(chrm.testConf).replace(", ", ",")+";;TRpaf2="+str(chrm.testFitness).replace(", ", ",")+"\n");
# retVal = None;
tempTime = Common.getCurrentTimeMil()
newPop = []
for i in range(self.sizeTopP):
newPop.append(pop[i])
i = 0
for i in range(0, len(pop) - self.sizeTopP, 2):
idx1 = 0
idx2 = 0
while (idx1 == idx2):
if (cnt >= 3):
idx1 = np.random.randint(len(pop))
idx2 = np.random.randint(len(pop))
else:
idx1 = GA.tornament(pop)
idx2 = GA.tornament(pop)
cnt += 1
cnt = 0
ds1 = pop[idx1].ds
ds2 = pop[idx2].ds
while True:
ds1, ds2 = GA.crossOver(ds1,
ds2,
fixedTrainSize,
isCount=self.isCount)
if len(set(list(ds1[:, -1]))) >= 2 and len(
set(list(ds2[:, -1]))) >= 2:
break
self.prnt('repeat cross')
while True:
ds1 = GA.Mutate(ds1, isCount=self.isCount, mad=mad)
if len(set(list(ds1[:, -1]))) >= 2:
break
self.prnt(
'repeat mut ds1, because all elements are of type one class'
)
while True:
ds2 = GA.Mutate(ds2, isCount=self.isCount, mad=mad)
if len(set(list(ds2[:, -1]))) >= 2:
break
self.prnt(
'repeat mut ds1, because all elements are of type one class'
)
if self.isCount:
newPop.append(
CHRM_GIS_Count(ds1, None, extraAsFitness='p-val'))
newPop.append(
CHRM_GIS_Count(ds2, None, extraAsFitness='p-val'))
else:
newPop.append(
CHRM_GIS(ds1, None, extraAsFitness='p-val'))
newPop.append(
CHRM_GIS(ds2, None, extraAsFitness='p-val'))
spentISTime += Common.getCurrentTimeMil() - tempTime
for i in range(len(newPop)):
tempTime = Common.getCurrentTimeMil()
pv, p_vals = DPLIB.checkSimilarity(newPop[i].ds[:, :-1],
testPart[:, :-1])
newPop[i].addToExtra('p-val', sum(p_vals))
spentISTime += Common.getCurrentTimeMil() - tempTime
tempTime = Common.getCurrentTimeMil()
newPop = DPLIB.MySort(newPop)
exit = False
countComp = 0
newPop, rdel = DPLIB.CombinePops(pop, newPop)
if (log):
pass
#retVal = ""
#for i in range(len(rdel)):
# chrm = rdel[i];
# retVal = DPLIB.getStats(chrm.ds, False, False, False);
# self.prnt("#POPDELITNFO;;gn="+str(g)+";;prt="+str(p)+";;For="+name+"@"+":"+retVal+";;rpaf="+str(chrm.fitness).replace(", ", ",")
# +";;conf="+str(chrm.conf).replace(", ", ",")+";;fit="+str(chrm.getFitness())+";;TConf2="+str(chrm.testConf).replace(", ", ",")+";;TRpaf2="+str(chrm.testFitness).replace(", ", ",")
# +"\n");
# retVal = None;
rdel = None
diff = abs(
GA.GetMeanFittness(pop, countComp) -
GA.GetMeanFittness(newPop, countComp))
if (diff < 0.000001):
exit = True
diffp.append(diff)
pop = newPop
if (pop[0].getFitness() > 0.0) and (exit):
break
exit = False
spentISTime += Common.getCurrentTimeMil() - tempTime
w = []
if (self.count == 0):
self.count = len(pop)
for i in range(self.count):
l = GLOB(clfName).getClassifier()
tds = pop[i].ds
self.FinalLearners.append(l)
self.FinalDatasets.append(tds)
testPartI = testPart
l.buildClassifier(tds)
if self.isCount:
actual = DPLIB.getActuals(testPartI)
prr = l.evaluateModel(testPartI)
#vals = DPLIB.getMeasuresCount(actual,prr)
actall = None
predall = None
if (len(actuals) == self.count):
actuals[i] = actuals[i] + actual
prrs[i] = prrs[i] + prr
else:
actuals.append(actual)
prrs.append(prr)
else:
vec = l.evaluateModel(testPartI)
if (len(preds) == self.count):
preds[i] += list(vec)
else:
preds.append(list(vec))
if (log):
pass
#retVal = DPLIB.getStats(tds, True, True, True);
#self.prnt("#TRPRTNFO;;prt="+str(p)+";;For="+name+"@"+":"+retVal+"\n");
#retVal = DPLIB.getStats(testPart,true,true, True);
#self.prnt("#TSTPRTNFO;;prt="+str(p)+";;For="+name+"@"+":"+retVal+"\n");
#vals = DPLIB.getConfMatrix(testPart[:,-1],vec)
#self.prnt("#TSTPRTVALS;;prt="+str(p)+";;For="+name+"@"+":"+
# "rpaf="+str(DPLIB.getMeasures(vals)).replace(", ", ",")
# +";;conf="+str(vals).replace(", ", ",")+"\n");
#retVal = None;
w.append(pop[i].getFitness())
isOK = True
if not isOK:
pass
else:
thresholds.append(pop[0].getFitness())
self.prnt()
self.prnt("Best Top Fitness:" + str(pop[0].fitness))
self.prnt("Best Fitness (mean):", pop[0].getMeanFitness())
if self.isCount:
vals = DPLIB.getMeasuresCountSet(actuals, prrs)
else:
vals1 = DPLIB.getConfMatrixSet(testSet[:, -1], preds)
vals = DPLIB.getMeasures(vals1)
if (isOK):
if not self.isCount:
if (len(preds) == 1):
auc = DPLIB.getAUC(testSet[:, -1], preds[0])
else:
auc = DPLIB.getAUCSet(testSet[:, -1], preds)
vals['auc'] = auc
self.prnt()
self.prnt("#CONF-TEST:" + name + ":" + self.file + ": " +
str(vals1))
self.prnt()
self.prnt(name + ":" + self.file + ": " + str(vals))
self.prnt()
else:
self.prnt()
self.prnt(name + ":" + self.file + ": " + str(vals))
self.prnt()
else:
bestI = pop[0]
rejectedFits.append(bestI.getFitness())
rejVals = copy.deepcopy(bestI.fitness)
rejectedPerfs.append(rejVals)
testRejVals = copy.deepcopy(vals)
rejectedTestPerfs.eppend(testRejVals)
self.prnt("#NOTOKPREDS----" + name + ":" + self.file + ": " +
str(vals))
if not self.isCount:
self.prnt()
self.prnt("#NOTOKPREDS----" + "#CONF-TEST:" + name + ":" +
self.file + ": " + str(vals1))
time = Common.getCurrentTimeMil() - startTime
self.prnt("#TIME-FOR:" + name + ":" + self.file + ": " + str(time))
self.prnt("#TIME-FOR-IS:" + name + ":" + self.file + ": " +
str(spentISTime))
return isOK