手写KL距离

logistic/analy_mse_entropy.py

@@ -0,0 +1,60 @@
+#!/usr/bin/python
+# -*- coding: UTF-8 -*-
+import sys
+reload(sys) 
+sys.setdefaultencoding('utf-8') 
+import random
+import math
+
+'''
+逻辑回归的mse即KL距离
+'''
+def logistic(x,w):
+	d=sum([ s1*s2 for [s1,s2] in zip(x,w)])
+	r=1.0/(1+math.exp(-1*d))
+	return r
+
+def cal_entropy(s1,s2):
+	s2=min(0.999,max(s2,0.001))
+	if s1==0:
+		return math.log(1.0/(1-s2))
+	else:
+		return math.log(1.0/s2)
+
+def cal_error(data,w):
+	y2=[ logistic(x,w) for [x,_] in data]
+	y=[s[1][0] for s in data]
+	mse=sum([   (s1-s2)*(s1-s2) for [s1,s2] in zip(y,y2)])/len(data)
+	entropy=sum([   cal_entropy(s1,s2) for [s1,s2] in zip(y,y2)])/len(data)
+	return  mse,entropy
+
+
+def read_data(path):
+	with open(path) as f :
+		lines=f.readlines()
+	lines=[eval(line.strip()) for line in lines]
+	return lines
+
+data=read_data("train_data")
+results=[]
+num=100
+step=0.2
+for i in range(-num,num):
+	w1=1.87+step*i
+	#for j in range(-num,num):
+	w2=-1.87#+step*j
+	e1,e2=cal_error(data,[w1,w2])
+	results.append("{},{},{}".format(w1,e1,e2))
+	print "{},{},{}".format(w1,e1,e2)
+
+# with open("mse_entropy.csv","w") as f  :
+# 	f.writelines("\n".join(results))
+
+
+
+
+
+
+
+
+