TensorFlow 学习 - MNIST 之CNN分类

在 TensorFlow 学习 - MNIST 之 SoftMax Regression 实现（完整代码，拷贝可运行） 中详细讲解了利用TensorFlow 对MNIST数据集分类的过程，精度在91%左右。这一篇利用卷积神经网络改善模型。

同样的，本文先附上完整可运行的代码，再分段介绍。

from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)

import tensorflow as tf
def weight_v(shape):
  initial = tf.truncated_normal(shape, stddev=0.1)
  return tf.Variable(initial)

def bias_v(shape):
  initial = tf.constant(0.1, shape=shape)
  return tf.Variable(initial)

def conv2d(x, W):
  return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

def max_pool_2x2(x):
  return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                        strides=[1, 2, 2, 1], padding='SAME')

sess = tf.InteractiveSession()

x = tf.placeholder(tf.float32, [None,784])
y_ = tf.placeholder("float",[None,10])
x_image = tf.reshape(x, [-1,28,28,1])

# first level

W_conv1 = weight_v([5, 5, 1, 32])
b_conv1 = bias_v([32])

h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)

# second level
W_conv2 = weight_v([5, 5, 32, 64])
b_conv2 = bias_v([64])

h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)

# line
W_fc1 = weight_v([7 * 7 * 64, 1024])
b_fc1 = bias_v([1024])

h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

#dropout
keep_prob = tf.placeholder("float")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

#output
W_fc2 = weight_v([1024, 10])
b_fc2 = bias_v([10])

y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)


cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
sess.run(tf.initialize_all_variables())
for i in range(20000):
  batch = mnist.train.next_batch(50)
  if i%100 == 0:
    train_accuracy = accuracy.eval(feed_dict={
        x:batch[0], y_: batch[1], keep_prob: 1.0})
    print "step %d, training accuracy %g"%(i, train_accuracy)
  train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})

print "test accuracy %g"%accuracy.eval(feed_dict={
    x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0})

定义函数

def weight_v(shape):
  initial = tf.truncated_normal(shape, stddev=0.1)
  return tf.Variable(initial)

def bias_v(shape):
  initial = tf.constant(0.1, shape=shape)
  return tf.Variable(initial)

def conv2d(x, W):
  return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

def max_pool_2x2(x):
  return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                        strides=[1, 2, 2, 1], padding='SAME')

为方便起见，将shape等操作定义为函数。 分别代表权重，偏置，卷积（卷积的strides为1），池化（2 * 2） 的filter池化。

初始化操作

sess = tf.InteractiveSession()
x = tf.placeholder(tf.float32, [None,784])
y_ = tf.placeholder("float",[None,10])
x_image = tf.reshape(x, [-1,28,28,1])

x 为原始的输入，y_ 为真实的标签，将初始的x（一个图像为例） reshape为 28 * 28 * 1 的图像。

卷积层

W_conv1 = weight_v([5, 5, 1, 32])
b_conv1 = bias_v([32])

定义5*5 * 1（一个图像为例）的filter，指定32个特征。因为设置的padding = ‘SAME’, 卷积之后的维度不变。依旧是28 * 28 * 1。

激活函数:

$$y = relu(W\times x + b)$$

h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)

池化层

h_pool1 = max_pool_2x2(h_conv1)

28 * 28 * 1 经由2 * 2 的filter 得到 14 * 14 * 1 的特征，此处选择max_pool。解释见 卷积神经网络 - 卷积池化

重复 卷积+ 激活+ 池化

W_conv2 = weight_v([5, 5, 32, 64])
b_conv2 = bias_v([64])

h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)

14 * 14 * 1 的特征经由 ２* 2 的filter 得到 7 * 7 * 1 的特征

全连接

W_fc1 = weight_v([7 * 7 * 64, 1024])
b_fc1 = bias_v([1024])

h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

加入一个有1024个神经元的全连接层处理整个图片

Dropout

keep_prob = tf.placeholder("float")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

为了减少过拟合，在输入层之前加入dropout，输出在dropout中保持不变的概率。

输出层

W_fc2 = weight_v([1024, 10])
b_fc2 = bias_v([10])

y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)

利用softmax函数判别

训练评估
详细见 TensorFlow 学习 - MNIST 之 SoftMax Regression 实现（完整代码，拷贝可运行）

cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
sess.run(tf.initialize_all_variables())
for i in range(20000):
  batch = mnist.train.next_batch(50)
  if i%100 == 0:
    train_accuracy = accuracy.eval(feed_dict={
        x:batch[0], y_: batch[1], keep_prob: 1.0})
    print "step %d, training accuracy %g"%(i, train_accuracy)
  train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})

print "test accuracy %g"%accuracy.eval(feed_dict={
    x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0})

结果输出
（part）