机器学习经典算法之(二十三) 随机梯度下降法

(一)  随机梯度下降法:

    上一篇博文,已经介绍了梯度下降算法。在实际中,为了实现梯度下降,往往分为随机梯度下降法和批量梯度下降法。

    随机梯度下降法基本思路:

    for j in range(j):

    对于每一次更新参数,不必遍历所有的训练集合,仅仅使用了一个数据,来变换一个参数。这样做不如完全梯度下降的精确度高,可能会走很多弯路,但整体趋势是走向minmum。

这样做可以节省更多的时间,算法更快。

   (二)  实现原理(参考《机器学习实战》)

from numpy import *

def loadDataSet():

   dataMat = []; labelMat = []

   fr = open('testSet.txt')

   for line in fr.readlines():

       lineArr = line.strip().split()

       dataMat.append([1.0, float(lineArr[0]), float(lineArr[1])])

       labelMat.append(int(lineArr[2]))

   return dataMat,labelMat

def stocGradAscent0(dataMatrix,classLabels):

   m,n = shape(dataMatrix)

   alpha = 0.01

   weights = ones(n)   #initialize toall ones

   for i in range(m):

       h = sigmoid(sum(dataMatrix[i]*weights))

       error = classLabels[i] - h

       weights = weights + alpha * error * dataMatrix[i]

   return weights

 

def stocGradAscent1(dataMatrix,classLabels, numIter=150):

   m,n = shape(dataMatrix)

   weights = ones(n)   #initialize toall ones

   for j in range(numIter):

       dataIndex = range(m)

       for i in range(m):

           alpha = 4/(1.0+j+i)+0.0001   

#apha decreases with iteration, does not

           randIndex = int(random.uniform(0,len(dataIndex)))

#go to 0 because of the constant

           h = sigmoid(sum(dataMatrix[randIndex]*weights))

           error = classLabels[randIndex] - h

           weights = weights + alpha * error * dataMatrix[randIndex]

           del(dataIndex[randIndex])

return weights

(三)、sklearn中随机梯度下降法应用举例:

from sklearn.linear_model importSGDClassifier

help(SGDClassifier)

clf = SGDClassifier(loss="hinge",penalty="l2")

from sklearn.linear_model importSGDClassifier

X = [[0., 0.], [1., 1.]]

y = [0, 1]

clf = SGDClassifier(loss="hinge",penalty="l2")

'''

SGDClassifier(alpha=0.0001, average=False,class_weight=None, epsilon=0.1,eta0=0.0, fit_intercept=True, l1_ratio=0.15,

       learning_rate='optimal', loss='hinge',max_iter=None, n_iter=None,n_jobs=1, penalty='l2', power_t=0.5,random_state=None,

      shuffle=True, tol=None, verbose=0, warm_start=False)

'''

clf.fit(X, y)

clf.predict([[1., .8]])

clf.coef_

clf.intercept_

clf.decision_function([[2., 2.]])

clf =SGDClassifier(loss="log",penalty="l2").fit(X, y)

print("predict_proba",clf.predict_proba([[3.,3.]]))

clf =SGDClassifier(loss="log",penalty="elasticnet").fit(X, y)

clf.predict_proba([[3., 3.]])

你可能感兴趣的:(机器之心,修炼之路)