1 import os
  2 import tab
  3 import tensorflow as tf
  4 
  5 print "tensorflow 5.2 "
  6 
  7 from tensorflow.examples.tutorials.mnist import input_data
  8 
  9 '''
 10 mnist = input_data.read_data_sets("/asky/tensorflow/mnist_data",one_hot=True)
 11 print "-------------------------------------"
 12
 
 print "Training data size: ",mnist.train.num_examples
 13 print "-------------------------------------"
 14 print "Validating data size: ",mnist.validation.num_examples
 15 print "-------------------------------------"
 16 print "Testing data size: " ,mnist.test.num_examples
 17 print "-------------------------------------"
 
 18 print "Example training data: ",mnist.train.images[0]
 19 print "-------------------------------------"
 20 print "Example training data label: ",mnist.train.labels[0]
 21 
 22 batch_size = 100
 23 xs,ys=mnist.train.next_batch(batch_size)
 24 
 25 print "X shape:",xs.shape
 26 
 27 print "Y shape:",ys.shape
 
 28 
 29 
 30 print "Test Tezt"
 31 '''
 32 
 33 INPUT_NODE = 784
 34 OUTPUT_NODE = 10
 35 
 36 LAYER1_NODE = 500
 37 
 38 BATCH_SIZE = 100
 39 
 40 LEARNING_RATE_BASE = 0.8
 41 LEARNING_RATE_DECAY = 0.99
 42 
 43 REGULARIZATION_RATE = 0.0001
 44 TRAINING_STEPS = 30000
 45 MOVING_AVERAGE_DECAY = 0.99
 46 
 47 def inference(input_tensor,avg_class,weights1,biases1,weights2,biases2):
 48     if avg_class == None:
 49         layer1 = tf.nn.relu(tf.matmul(input_tensor,weights1)+biases1)
 50         return tf.matmul(layer1,weights2)+biases2
 51     else:
 52         layer1 = tf.nn.relu(
 53             tf.matmul(input_tensor,avg_class.average(weights1))+
 54             avg_class.average(biases1))
 55         return tf.matmul(layer1,avg_class.average(weights2))+avg_class.average(biases2)
 56 
 57 def train(mnist):
 58     x = tf.placeholder(tf.float32,[None,INPUT_NODE],name='x-input')
 59     y_ = tf.placeholder(tf.float32,[None,OUTPUT_NODE],name='y-input')
 60     weights1 = tf.Variable(
 61         tf.truncated_normal([INPUT_NODE,LAYER1_NODE],stddev=0.1))
 62     biases1 = tf.Variable( tf.constant(0.1,shape=[LAYER1_NODE]))
 63 
 64     weights2 = tf.Variable(tf.truncated_normal([LAYER1_NODE,OUTPUT_NODE],stddev=0.1))
 65     biases2 = tf.Variable(tf.constant(0.1,shape=[OUTPUT_NODE]))
 66 
 67     y = inference(x,None,weights1,biases1,weights2,biases2)
 68 
 69     global_step = tf.Variable(0,trainable=False)
 70 
 71     variable_averages = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY,global_step)
 72 
 73     variables_averages_op = variable_averages.apply(tf.trainable_variables())
 74 
 75     average_y = inference(x,variable_averages,weights1,biases1,weights2,biases2)
 76 
 77     #cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(y, tf.argmax(y_, 1 ))
 78     cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=tf.argmax(y_, 1), logits=y)
 79 
 80     cross_entropy_mean = tf.reduce_mean(cross_entropy)
 81 
 82     regularizer = tf.contrib.layers.l2_regularizer(REGULARIZATION_RATE)
 83 
 84     regularization = regularizer(weights1) + regularizer(weights2)
 85 
 86     loss = cross_entropy_mean + regularization
 87 
 88     learning_rate = tf.train.exponential_decay(
 89         LEARNING_RATE_BASE,
 90         global_step,
 91         mnist.train.num_examples/BATCH_SIZE,
 92         LEARNING_RATE_DECAY
 93     )
 94 
 95     train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)
 96 
 97     with tf.control_dependencies([train_step,variables_averages_op]):
 98         train_op = tf.no_op(name='train')
 99 
100     correct_prediction = tf.equal(tf.argmax(average_y,1),tf.argmax(y_,1))
101     accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
102 
103     with tf.Session() as sess:
104         tf.global_variables_initializer().run()
105         validate_feed = {x: mnist.validation.images,
106                          y_: mnist.validation.labels}
107         test_feed = {x: mnist.test.images, y_: mnist.test.labels }
108         for i in range(TRAINING_STEPS):
109             if i % 1000 ==0:
110                 validate_acc = sess.run(accuracy,feed_dict=validate_feed)
111                 print ("After %d training step(s),validation accuracy "
112                         "using average model is %g " %(i,validate_acc) )
113             xs, ys = mnist.train.next_batch(BATCH_SIZE)
114             sess.run(train_op,feed_dict={x: xs , y_ : ys})
115 
116         test_acc  = sess.run(accuracy,feed_dict=test_feed)
117         print ( "After %d training step(s),test accuracy using average "
118                 "model is %g " % (TRAINING_STEPS , test_acc)  )
119 
120 def main(argv=None) :
121     mnist = input_data.read_data_sets("/asky/tensorflow/mnist_data",one_hot=True)
122     train(mnist)
123 
124 if __name__ == '__main__':
125     tf.app.run()