基于Boost方法的人脸检测（4）：弱分类器生成、强分类器生成（AdaBoost）

弱分类器生成：

allWeakCLF=[]
for feaInd in range(data.shape[1]-1):
	tempD=data.iloc[:,[feaInd,-1]]
	tempD=tempD.sort(columns=tempD.columns[0], ascending=False)
	tempD=np.array(tempD)
	
	bestCLF=[0.0, 100000, 0] #threshold, errorNum, 0 or 1(flag)
	#0: errorNum is based on ">= threshold is positive"
	#1: errorNum is based on ">= threshold is negative"
	curPosNum=0
	curNegNum=0
	for threshold, label in tempD:
		if label==1:
			curPosNum+=1
		else:
			curNegNum+=1
		errorNum0=(posNum-curPosNum)+curNegNum
		errorNum1=(negNum-curNegNum)+curPosNum
		if errorNum0>=errorNum1:
			if errorNum1<bestCLF[1]:
				bestCLF=[threshold, errorNum1, 1] #threshold, errorNum, 0 or 1(flag)
		else: #errorNum0<errorNum1
			if errorNum0<bestCLF[1]:
				bestCLF=[threshold, errorNum0, 0] #threshold, errorNum, 0 or 1(flag)
	allWeakCLF.append(bestCLF)
joblib.dump(allWeakCLF, 'allWeakCLF.pkl')
#allWeakCLF=joblib.load('allWeakCLF.pkl')

随机方式生成强分类器或固定方式生成强分类器：

def MyAdaBoost(data, WeakClfInfo):
	
	#1: init
	dataWeight=[1.0/len(data)]*len(data)
	
	#2: generate strong clf
	#errorNum, ind, threshold, 0 or 1(flag), clfWeight(init with 1)
	for j, (errorNum, ind, threshold, flag, clfWeight) in enumerate(WeakClfInfo): #for each weak clf
		
		#2.1: caculate normalized weighted error(delta)
		errorFlag=[0]*len(data) #0:correct, 1:incorrect
		errorWeight=0.0
		totalWeight=0.0
		if flag==0: #0: errorNum is based on ">= threshold is positive"
			for i, (value, label) in enumerate(data[:,[ind, -1]]): #for each data
				totalWeight+=dataWeight[i]
				if value>=threshold and label==0:
					errorWeight+=dataWeight[i]
					errorFlag[i]=1
				elif value<threshold and label==1:
					errorWeight+=dataWeight[i]
					errorFlag[i]=1
		else: #flag==1, #1: errorNum is based on ">= threshold is negative"
			for i, (value, label) in enumerate(data[:,[ind, -1]]): #for each data
				totalWeight+=dataWeight[i]
				if value>=threshold and label==1:
					errorWeight+=dataWeight[i]
					errorFlag[i]=1
				elif value<threshold and label==0:
					errorWeight+=dataWeight[i]
					errorFlag[i]=1
		delta=errorWeight/totalWeight
		delta=math.sqrt((1-delta)/delta)
		
		#2.2: update dataWeight according to delta
		for i, flag in enumerate(errorFlag):
			if flag==0: #0:correct, 1:incorrect
				dataWeight[i]/=delta
			else:
				dataWeight[i]*=delta
		
		#2.3: update weak clf weight(and errorNum) according to delta
		WeakClfInfo[j][0]=delta #errorNum ==> normalized weighted error(delta)
		WeakClfInfo[j][4]=math.log(delta) #clfWeight
		#not WeakClfInfo[0]!!!!!! take a long time to find this bug!!!!
		
	#3: return the strong clf(WeakClfInfo now has the strong info)
	return WeakClfInfo


#how many strClf we want to train, weakClfNum in each strClf
def RandomGenerateStrClf(data, allWeakCLF, strClfNum=10, weakClfNum=60):
	randomStrClf=[]
	weakClfInd=range(len(allWeakCLF)) #weakClfInd~=feaInd
	for i in range(strClfNum):
		print "training the", i+1, "strong classifier", "^_^"*10
		sampledWeakClfInd=random.sample(weakClfInd,weakClfNum)
		sampledWeakClfInfo=[] #errorNum, ind, threshold, 0 or 1(flag), clfWeight(init with 1)
		for ind in sampledWeakClfInd:
			sampledWeakClfInfo.append([allWeakCLF[ind][1], ind, allWeakCLF[ind][0], allWeakCLF[ind][2], 1])
		sortedSampledWeakClfInfo=sorted(sampledWeakClfInfo)
		strongClfInfo=MyAdaBoost(data, sortedSampledWeakClfInfo)
		#strongClfInfo ==> list of [errorNum->delta, ind, threshold, 0 or 1(flag), new_clfWeight]
		randomStrClf.append(strongClfInfo)
	return randomStrClf

#how many strClf we want to train, weakClfNum in each strClf
def StaticGenerateStrClf(data, sortedWeakClfInfo, strClfNum=10, weakClfNum=20):
	staticStrClf=[]
	step=len(allWeakCLF)/strClfNum
	if step>weakClfNum:
		step=weakClfNum
	for i in range(strClfNum):
		print "training the", i+1, "strong classifier", "^_^"*10
		strongClfInfo=MyAdaBoost(data, sortedWeakClfInfo[i*step:i*step+weakClfNum])
		#strongClfInfo ==> list of [errorNum->delta, ind, threshold, 0 or 1(flag), new_clfWeight]
		staticStrClf.append(strongClfInfo)
	return staticStrClf