TensorFlow是目前深度学习最流行的框架，很有学习的必要，下面我们就来实际动手，使用TensorFlow搭建一个简单的CNN，来对经典的mnist数据集进行数字识别。

如果对CNN还不是很熟悉的朋友，可以参考：Convolutional Neural Network。

下面就开始。

step 0 导入TensorFlow

1 import tensorflow as tf
2 from tensorflow.examples.tutorials.mnist import input_data

step 1 加载数据集mnist

声明两个placeholder，用于存储神经网络的输入，输入包括image和label。这里加载的image是(784,)的shape。

1 mnist = input_data.read_data_sets('MNIST_data/', one_hot=True)
2 x = tf.placeholder(tf.float32,[None, 784])
3 y_ = tf.placeholder(tf.float32, [None, 10])

step 2 定义weights和bias

为了使代码整洁，这里把weight和bias的初始化封装成函数。

1 #----Weight Initialization---#
2 #One should generally initialize weights with a small amount of noise for symmetry breaking, and to prevent 0 gradients
3 def weight_variable(shape):
4     initial = tf.truncated_normal(shape, stddev=0.1)
5     return tf.Variable(initial)
6 def bias_variable(shape):
7     initial = tf.constant(0.1, shape=shape)
8     return tf.Variable(initial)

step 3 定义卷积层和maxpooling

同样，为了代码的整洁，将卷积层和maxpooling封装起来。padding=‘SAME’表示使用padding，不改变图片的大小。

1 #Convolution and Pooling
2 #Our convolutions uses a stride of one and are zero padded so that the output is the same size as the input.
3 #Our pooling is plain old max pooling over 2x2 blocks
4 def conv2d(x, W):
5     return tf.nn.conv2d(x, W, strides=[1,1,1,1], padding='SAME')
6 def max_pool_2x2(x):
7     return tf.nn.max_pool(x, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')

step 4 reshape image数据

为了神经网络的layer可以使用image数据，我们要将其转化成4d的tensor: (Number, width, height, channels)

1 #To apply the layer, we first reshape x to a 4d tensor, with the second and third dimensions corresponding to image width and height,
2 #and the final dimension corresponding to the number of color channels.
3 x_image = tf.reshape(x, [-1,28,28,1])

下面我们就要开始搭建CNN结构了。

step 5 搭建第一个卷积层

使用32个5x5的filter，然后通过maxpooling。

 1 #----first convolution layer----#
 2 #he convolution will compute 32 features for each 5x5 patch. Its weight tensor will have a shape of [5, 5, 1, 32].
 3 #The first two dimensions are the patch size,
 4 #the next is the number of input channels, and the last is the number of output channels.
 5 W_conv1 = weight_variable([5,5,1,32])
 6 
 7 #We will also have a bias vector with a component for each output channel.
 8 b_conv1 = bias_variable([32])
 9 
10 #We then convolve x_image with the weight tensor, add the bias, apply the ReLU function, and finally max pool.
11 #The max_pool_2x2 method will reduce the image size to 14x14.
12 h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
13 h_pool1 = max_pool_2x2(h_conv1)

step 6 第二层卷积

使用64个5x5的filter。

1 #----second convolution layer----#
2 #The second layer will have 64 features for each 5x5 patch and input size 32.
3 W_conv2 = weight_variable([5,5,32,64])
4 b_conv2 = bias_variable([64])
5 
6 h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
7 h_pool2 = max_pool_2x2(h_conv2)

step 7 构建全链接层

需要将上一层的输出，展开成1d的神经层。

1 #----fully connected layer----#
2 #Now that the image size has been reduced to 7x7, we add a fully-connected layer with 1024 neurons to allow processing on the entire image
3 W_fc1 = weight_variable([7*7*64, 1024])
4 b_fc1 = bias_variable([1024])
5 
6 h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
7 h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat,W_fc1) + b_fc1)

step 8 添加Dropout

加入Dropout层，可以防止过拟合问题。注意，这里使用了另外一个placeholder，可以控制在训练和预测时是否使用Dropout。

1 #-----dropout------#
2 #To reduce overfitting, we will apply dropout before the readout layer.
3 #We create a placeholder for the probability that a neuron's output is kept during dropout.
4 #This allows us to turn dropout on during training, and turn it off during testing.
5 keep_prob = tf.placeholder(tf.float32)
6 h_fc1_dropout = tf.nn.dropout(h_fc1, keep_prob)

step 9 输入层

没有什么特别的，就是输出一个线性结果。

1 #----read out layer----#
2 W_fc2 = weight_variable([1024,10])
3 b_fc2 = bias_variable([10])
4 y_conv = tf.matmul(h_fc1_dropout, W_fc2) + b_fc2

step 10 训练和评估

首先，需要指定一个cost function --cross_entropy，在输出层使用softmax。然后指定optimizer--adam。需要特别指出的是，一定要记得

tf.global_variables_initializer().run()初始化变量

 1 #------train and evaluate----#
 2 cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_conv))
 3 train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
 4 accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(y_, 1), tf.argmax(y_conv, 1)), tf.float32))
 5 with tf.Session() as sess:
 6     tf.global_variables_initializer().run()
 7     for i in range(3000):
 8         batch = mnist.train.next_batch(50)
 9         if i % 100 == 0:
10             train_accuracy = accuracy.eval(feed_dict = {x: batch[0],
11                                                        y_: batch[1],
12                                                        keep_prob: 1.})
13             print('setp {},the train accuracy: {}'.format(i, train_accuracy))
14         train_step.run(feed_dict = {x: batch[0], y_: batch[1], keep_prob: 0.5})
15     test_accuracy = accuracy.eval(feed_dict = {x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.})
16     print('the test accuracy :{}'.format(test_accuracy))
17     saver = tf.train.Saver()
18     path = saver.save(sess, './my_net/mnist_deep.ckpt')
19     print('save path: {}'.format(path))