lr

#encoding=gbk 
import random 
import math

#读取数据集
def load_data(filename):
	dataset = file(filename)
	x_input=[]
	y_output=[]
	for line in dataset:
		a_x_input=[]
		splitted_line = line.strip().split(',')
		length=len(splitted_line)
		a_x_input=splitted_line[0:length-1]
		a_x_input.append(1)
		x_input.append(a_x_input)
		y_output.append(splitted_line[length-1])

	return (x_input, y_output)

def sigmoid(x):
	return 1.0/(1.0+math.pow(math.e, -x))

def compute_log_likelihood(trainset_x, trainset_y,weights):
	sample_num=len(trainset_x)
	feature_num=len(trainset_x[0])
	log_likelihood=0.0
	for i in range(0,sample_num):
		wx=0
		for j in range(0, feature_num):
			wx+=weights[j]*float(trainset_x[i][j])

		log_likelihood+=(float(trainset_y[i])*wx-math.log(1+math.pow(math.e, wx)))

	print "log_likelihood", log_likelihood
		
	return log_likelihood

#梯度上升
def train(trainset_x, trainset_y, max_iter):
	sample_num=len(trainset_x)
	feature_num=len(trainset_x[0])
	weights=[1]*feature_num

	alpha=0.001
	print "sample_num, feature_num", sample_num, feature_num

	for m in range(max_iter): #0到max_iter-1
		old_weights=weights[0:len(weights)]

		old_log_likelihood = compute_log_likelihood(trainset_x, trainset_y, old_weights)

		error=[0]*sample_num
		for i in range(0,sample_num): #0到sample_num-1
			tmp=0	
			for j in range(0, feature_num):
				tmp+=float(trainset_x[i][j])*old_weights[j]
			output=sigmoid(tmp)
			error[i]+=float(trainset_y[i])-output

		for k in range(0, feature_num): #0到feature_num-1
			gradient=0
			for i in range(0, sample_num):
				gradient+=error[i]*float(trainset_x[i][k])
			weights[k]=old_weights[k]+alpha*gradient

		new_log_likelihood = compute_log_likelihood(trainset_x, trainset_y, weights)

		if new_log_likelihood < old_log_likelihood :
			print "new", new_log_likelihood
			print "old", old_log_likelihood
			print "error"
			print (new_log_likelihood-old_log_likelihood)/old_log_likelihood
			#break
		if abs((new_log_likelihood-old_log_likelihood)/old_log_likelihood) < 0.001 :
			print (new_log_likelihood-old_log_likelihood)/old_log_likelihood
			#break
		old_log_likelihood=new_log_likelihood

	return weights

if __name__ == "__main__":
	dataset_file="ds1.10.csv"
	x_input, y_output = load_data(dataset_file)
	print len(x_input)

	trainset_x=[]
	trainset_y=[]

	testset_x=[]
	testset_y=[]
	max_iter=500

	for i in range(0,len(x_input) ):
		r = random.randint(1,10)  #产生训练集和测试集
		if r >2 :
			trainset_x.append(x_input[i])
			trainset_y.append(y_output[i])
		else:
			testset_x.append(x_input[i])
			testset_y.append(y_output[i])

	print len(trainset_x),len(trainset_y)
	print len(testset_x), len(testset_y)

	weights = train(trainset_x,trainset_y,max_iter)