Softmax

python 实现softmax

1、实验使用的数据来自http://sci2s.ugr.es/keel/category.php?cat=clas
2、有关具体的推导请参考Andrew Ng.的课程http://cs229.stanford.edu/

#-*- coding=utf-8 -*-import numpy as npclass Softmax(object):    "class softmax"    def __init__(self,size,epo=1000,rate=0.001):        """        :param size: size=(num of attributes,num of classes)        :param epo: training        :param rate:learning rate        :return:        """        self.epo=epo        self.rate=rate        self.weights=np.random.normal(size=size)    def fit(self,traindata,testdata=None):        "train softmax and use stochastic gradient decent updating weights"        best_accuracy=-1        best_epo=-1        for i in range(self.epo):            print "epo %d"%i            for j in range(data.shape[0]):                x=data[j,:-1]                y=data[j,-1]                h=self.softmax(x)                #print h                #update weights                self.weights=self.weights.transpose()                for k in range(self.weights.shape[0]):                    if k==y:                        self.weights[k]+=self.rate*(1-h[k])*x                    else:                        self.weights[k]+=self.rate*(-h[k])*x                self.weights=self.weights.transpose()            if traindata is not None:                accu=self.accuracy(traindata)                if accu>best_accuracy:                    best_epo=i                    best_accuracy=accu            print "best_epo is %d ,best_accuracy is %lf"%(best_epo,best_accuracy)    def softmax(self,x):        temp=self.weights.transpose().dot(x)        max=np.max(temp)        temp=temp-max        h=np.exp(temp)/np.sum(np.exp(temp))        h=np.nan_to_num(h)        return h    def predict(self,x):        h=np.argmax(self.softmax(x))        return h    def accuracy(self,data):        num=0        for i in range(data.shape[0]):            x=data[i][:-1]            y=data[i][-1]            h=self.predict(x)            if y==h:                num+=1        #print "predict accuracy is %lf"%(num*1.0/data.shape[0])        return num*1.0/data.shape[0]def loadData(path):    data=np.loadtxt(path,skiprows=21,dtype="int32",delimiter=",")    new_data=np.ones((data.shape[0],data.shape[1]+1))    """"relarge the dataset with x0=1"""    new_data[:,1:]=data    return new_datadata=loadData("D:\\SelfLearning\\Machine Learning\\ClassifyDataSet\\penbased\\penbased.dat")train_data=data[:8000]test_data=data[8000:]lr=Softmax(size=(17,10),epo=150,lamda=0.00001)lr.fit(train_data,test_data)print lr.accuracy(test_data)print lr.accuracy(train_data)

针对penbased.dat数据集进行测试，该数据集为手写体数字的十分类问题。
实验结果：

best_epo is 14 ,best_accuracy is 0.910625