Softmax回归（使用tensorflow）

 1 # coding:utf8
 2 import numpy as np
 3 import cPickle
 4 import os
 5 import tensorflow as tf
 6 
 7 class SoftMax:
 8     def __init__(self,MAXT=30,step=0.0025):
 9         self.MAXT = MAXT
10         self.step = step
11         
12     def load_theta(self,datapath="data/softmax.pkl"):
13         self.theta = cPickle.load(open(datapath,‘rb‘))
14 
15     def process_train(self,data,label,typenum=10,batch_size=500):
16         batches =  data.shape[0] / batch_size
17         valuenum=data.shape[1]
18         if len(label.shape)==1:
19             label=self.reshape_data(label,typenum)
20         x = tf.placeholder("float", [None,valuenum])
21         theta = tf.Variable(tf.zeros([valuenum,typenum]))
22         y = tf.nn.softmax(tf.matmul(x,theta))
23         y_ = tf.placeholder("float", [None, typenum])
24         cross_entropy = -tf.reduce_sum(y_*tf.log(y))  #交叉熵
25         train_step = tf.train.GradientDescentOptimizer(self.step).minimize(cross_entropy)
26         init = tf.initialize_all_variables()
27         sess = tf.Session()
28         sess.run(init)
29         for epoch in range(self.MAXT):
30             cost_=[]
31             for index in xrange(batches):
32                 c_,_=sess.run([cross_entropy,train_step], feed_dict={ x: data[index * batch_size: (index + 1) * batch_size],
33                 y_: label[index * batch_size: (index + 1) * batch_size]})
34                 cost_.append(c_)
35             if epoch % 5 == 0:
36                 print(( ‘epoch %i, minibatch %i/%i,averange cost is %f‘) %
37                         (epoch,index + 1,batches,np.mean(cost_)))
38         self.theta=sess.run(theta)
39         if not os.path.exists(‘data/softmax.pkl‘):
40             f= open("data/softmax.pkl",‘wb‘)
41             cPickle.dump(self.theta,f)
42             f.close()
43         return self.theta
44 
45 
46     def process_test(self,data,label,typenum=10):
47         valuenum=data.shape[1]
48         if len(label.shape)==1:
49             label=self.reshape_data(label,typenum)
50         x = tf.placeholder("float", [None,valuenum])
51         theta = self.theta
52         y = tf.nn.softmax(tf.matmul(x,theta))
53         y_ = tf.placeholder("float", [None, typenum])
54         init = tf.initialize_all_variables()
55         sess = tf.Session()
56         sess.run(init)
57         correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
58         accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
59         print "Accuracy: ",sess.run(accuracy, feed_dict={x: data,y_: label})
60 
61     def h(self,x):
62         m = np.exp(np.dot(x,self.theta))
63         sump = np.sum(m,axis=1)
64         return m/sump
65 
66     def predict(self,x):
67         return np.argmax(self.h(x),axis=1)
68 
69     def reshape_data(self,label,typenum):
70         label_=[]
71         for yl_ in label:
72             tl_=np.zeros(typenum)
73             tl_[yl_]=1.0
74             label_.append(tl_)
75         return np.mat(label_)
76 
77 if __name__ == ‘__main__‘:
78     f = open(‘mnist.pkl‘, ‘rb‘)
79     training_data, validation_data, test_data = cPickle.load(f)
80     training_inputs = [np.reshape(x, 784) for x in training_data[0]]
81     data = np.array(training_inputs)
82     training_inputs = [np.reshape(x, 784) for x in validation_data[0]]
83     vdata = np.array(training_inputs)
84     f.close()
85 
86     softmax = SoftMax()
87     softmax.process_train(data,training_data[1])
88     softmax.process_test(vdata,validation_data[1])  #Accuracy:  0.9269
89     softmax.process_test(data,training_data[1])  #Accuracy:  0.92718