1.  import tensorflow as tf
2.  hello = tf.constant('Hello, TensorFlow!')
3.  sess = tf.Session()
4.  print sess.run(hello)

 　　首先，通过tf.constant创建一个常量，然后启动Tensorflow的Session，调用sess的run方法来启动整个graph。

　　接下来我们做下简单的数学的方法：

1.  import tensorflow as tf
2.  a = tf.constant(2)
3.  b = tf.constant(3)
4.  with tf.Session() as sess:
5.      print "a=2, b=3"
6.      print "Addition with constants: %i" % sess.run(a+b)
7.      print "Multiplication with constants: %i" % sess.run(a*b)
8.  # output
9.  a=2, b=3
10.  Addition with constants: 5
11.  Multiplication with constants: 6

 　　接下来用tensorflow的placeholder来定义变量做类似计算：

 　　placeholder的使用见https://www.tensorflow.org/versions/r0.8/api_docs/python/io_ops.html#placeholder

1.  import tensorflow as tf
2.  a = tf.placeholder(tf.int16)
3.  b = tf.placeholder(tf.int16)
4.  add = tf.add(a, b)
5.  mul = tf.mul(a, b)
6.  with tf.Session() as sess:
7.      # Run every operation with variable input
8.      print "Addition with variables: %i" % sess.run(add, feed_dict={a: 2, b: 3})
9.      print "Multiplication with variables: %i" % sess.run(mul, feed_dict={a: 2, b: 3})
10.  # output:
11.  Addition with variables: 5
12.  Multiplication with variables: 6
13.  matrix1 = tf.constant([[3., 3.]])
14.  matrix2 = tf.constant([[2.],[2.]])
15.  with tf.Session() as sess:
16.      result = sess.run(product)
17.      print result

 线性回归

　　以下代码来自GitHub - aymericdamien/TensorFlow-Examples: TensorFlow Tutorial and Examples for beginners,仅作学习用

1.  import tensorflow as tf
2.  import numpy
3.  import matplotlib.pyplot as plt
4.  rng = numpy.random
5. 
   
6.  # Parameters
7.  learning_rate = 0.01
8.  training_epochs = 2000
9.  display_step = 50
10. 
    
11.  # Training Data
12.  train_X = numpy.asarray([3.3,4.4,5.5,6.71,6.93,4.168,9.779,6.182,7.59,2.167,7.042,10.791,5.313,7.997,5.654,9.27,3.1])
13.  train_Y = numpy.asarray([1.7,2.76,2.09,3.19,1.694,1.573,3.366,2.596,2.53,1.221,2.827,3.465,1.65,2.904,2.42,2.94,1.3])
14.  n_samples = train_X.shape[0]
15. 
    
16.  # tf Graph Input
17.  X = tf.placeholder("float")
18.  Y = tf.placeholder("float")
19. 
    
20.  # Create Model
21. 
    
22.  # Set model weights
23.  W = tf.Variable(rng.randn(), name="weight")
24.  b = tf.Variable(rng.randn(), name="bias")
25. 
    
26.  # Construct a linear model
27.  activation = tf.add(tf.mul(X, W), b)
28. 
    
29.  # Minimize the squared errors
30.  cost = tf.reduce_sum(tf.pow(activation-Y, 2))/(2*n_samples) #L2 loss
31.  optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost) #Gradient descent
32. 
    
33.  # Initializing the variables
34.  init = tf.initialize_all_variables()
35. 
    
36.  # Launch the graph
37.  with tf.Session() as sess:
38.      sess.run(init)
39. 
    
40.      # Fit all training data
41.      for epoch in range(training_epochs):
42.          for (x, y) in zip(train_X, train_Y):
43.              sess.run(optimizer, feed_dict={X: x, Y: y})
44. 
    
45.          #Display logs per epoch step
46.          if epoch % display_step == 0:
47.              print "Epoch:", '%04d' % (epoch+1), "cost=", \
48.                  "{:.9f}".format(sess.run(cost, feed_dict={X: train_X, Y:train_Y})), \
49.                  "W=", sess.run(W), "b=", sess.run(b)
50. 
    
51.      print "Optimization Finished!"
52.      print "cost=", sess.run(cost, feed_dict={X: train_X, Y: train_Y}), \
53.            "W=", sess.run(W), "b=", sess.run(b)
54. 
    
55.      #Graphic display
56.      plt.plot(train_X, train_Y, 'ro', label='Original data')
57.      plt.plot(train_X, sess.run(W) * train_X + sess.run(b), label='Fitted line')
58.      plt.legend()
59.      plt.show()

60. 逻辑回归

    1.  import tensorflow as tf
    2.  # Import MINST data
    3.  from tensorflow.examples.tutorials.mnist import input_data
    4.  mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)
    5. 
       
    6.  # Parameters
    7.  learning_rate = 0.01
    8.  training_epochs = 25
    9.  batch_size = 100
    10.  display_step = 1
    11. 
        
    12.  # tf Graph Input
    13.  x = tf.placeholder(tf.float32, [None, 784]) # mnist data image of shape 28*28=784
    14.  y = tf.placeholder(tf.float32, [None, 10]) # 0-9 digits recognition => 10 classes
    15. 
        
    16.  # Set model weights
    17.  W = tf.Variable(tf.zeros([784, 10]))
    18.  b = tf.Variable(tf.zeros([10]))
    19. 
        
    20.  # Construct model
    21.  pred = tf.nn.softmax(tf.matmul(x, W) + b) # Softmax
    22. 
        
    23.  # Minimize error using cross entropy
    24.  cost = tf.reduce_mean(-tf.reduce_sum(y*tf.log(pred), reduction_indices=1))
    25.  # Gradient Descent
    26.  optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
    27. 
        
    28.  # Initializing the variables
    29.  init = tf.initialize_all_variables()
    30. 
        
    31.  # Launch the graph
    32.  with tf.Session() as sess:
    33.      sess.run(init)
    34. 
        
    35.      # Training cycle
    36.      for epoch in range(training_epochs):
    37.          avg_cost = 0.
    38.          total_batch = int(mnist.train.num_examples/batch_size)
    39.          # Loop over all batches
    40.          for i in range(total_batch):
    41.              batch_xs, batch_ys = mnist.train.next_batch(batch_size)
    42.              # Run optimization op (backprop) and cost op (to get loss value)
    43.              _, c = sess.run([optimizer, cost], feed_dict={x: batch_xs,
    44.                                                            y: batch_ys})
    45.              # Compute average loss
    46.              avg_cost += c / total_batch
    47.          # Display logs per epoch step
    48.          if (epoch+1) % display_step == 0:
    49.              print "Epoch:", '%04d' % (epoch+1), "cost=", "{:.9f}".format(avg_cost)
    50. 
        
    51.      print "Optimization Finished!"
    52. 
        
    53.      # Test model
    54.      correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
    55.      # Calculate accuracy
    56.      accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
    57.      print "Accuracy:", accuracy.eval({x: mnist.test.images, y: mnist.test.labels})
    58. 
        
    59.      # result :
    60.      Epoch: 0001 cost= 29.860467369
    61.      Epoch: 0002 cost= 22.001451784
    62.      Epoch: 0003 cost= 21.019925554
    63.      Epoch: 0004 cost= 20.561320320
    64.      Epoch: 0005 cost= 20.109135756
    65.      Epoch: 0006 cost= 19.927862290
    66.      Epoch: 0007 cost= 19.548687116
    67.      Epoch: 0008 cost= 19.429119071
    68.      Epoch: 0009 cost= 19.397068211
    69.      Epoch: 0010 cost= 19.180813479
    70.      Epoch: 0011 cost= 19.026808132
    71.      Epoch: 0012 cost= 19.057875510
    72.      Epoch: 0013 cost= 19.009575057
    73.      Epoch: 0014 cost= 18.873240641
    74.      Epoch: 0015 cost= 18.718575359
    75.      Epoch: 0016 cost= 18.718761925
    76.      Epoch: 0017 cost= 18.673640560
    77.      Epoch: 0018 cost= 18.562128253
    78.      Epoch: 0019 cost= 18.458205289
    79.      Epoch: 0020 cost= 18.538211225
    80.      Epoch: 0021 cost= 18.443384213
    81.      Epoch: 0022 cost= 18.428727668
    82.      Epoch: 0023 cost= 18.304270616
    83.      Epoch: 0024 cost= 18.323529782
    84.      Epoch: 0025 cost= 18.247192113
    85.      Optimization Finished!
    86.      (10000, 784)
    87.      Accuracy 0.9206
    88. 
        
    89. 
        
    90. 
        
    91. 
        
    92. 
        
    93. 
        
    94. 
        
    95. 这里有个小插曲，ipython notebook在一个notebook打开时，一直在占用GPU资源，可能是之前有一个notebook一直打开着，然后占用着GPU资源，然后在计算Accuracy的”InternalError: Dst tensor is not initialized.” 然后找了github上面也有这个问题InternalError: Dst tensor is not initialized.,可以肯定是GPU的memory相关的问题，所以就尝试加上tf.device(‘/cpu:0’)，将Accuracy这步拉到cpu上计算，但是又出现OOM的问题，最后nvidia-smi时，发现有一个python脚本一直占用3g多的显存，把它kill之后恢复了，之前还比较吐槽怎么可能10000*784个float就把显存撑爆呢，原来是自己的问题。
        

    96.  

        　　这里逻辑回归，model是一个softmax函数用来做多元分类，大概意思是选择10当中最后预测概率最高作为最终的分类。

         

        　　其实基本的tensorflow没有特别好讲的，语法的课程什么可以去看看基本的文档，之后我会找一点经典有趣的tensorflow的代码应用来看看，毕竟『show me the code 』才是程序猿应有的态度。

         

        　　本文由『UCloud内核与虚拟化研发团队』提供。