TensorFlow-CNN

CNN

import tensorflow as tf

from tensorflow.examples.tutorials.mnist import input_data

mnist = input_data.read_data_sets('Mnist_data', one_hot = True)

def compute_accuracy(v_xs, v_ys):
    global prediction
    y_pre = sess.run(prediction, feed_dict = {xs : v_xs, keep_prob : 1})
    correct_prediction = tf.equal(tf.argmax(y_pre, 1), tf.argmax(v_ys, 1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
    result = sess.run(accuracy, feed_dict = {xs : v_xs, ys : v_ys, keep_prob : 1})
    return result

# define weight
def weight_variable(shape):
    #truncated_normal:截短正态分布，stddev：标准差
    initial = tf.truncated_normal(shape, stddev = 0.1)
    return tf.Variable(initial)

# define biases
def bias_variable(shape):
    initial = tf.constant(0.1, shape = shape)
    return tf.Variable(initial)

# cov2d 步长 [1, 1,]
def conv2d(x, W):
    #SAME：最终抽取的图像跟原图像一样(越界抽取)
    #strides (batch_size,x,y,通道)
    return tf.nn.conv2d(x, W, strides = [1, 1, 1, 1], padding = 'SAME')

# max_pool
def max_poo_2x2(x):
    #ksize:池化窗口的大小[1, height, width, 1],想在batch和channels上做池化，所以这两个维度设为了1

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

# define placeholder for inputting data
xs = tf.placeholder(tf.float32, [None, 784])
ys = tf.placeholder(tf.float32, [None, 10])

# define placeholder for dropout
keep_prob = tf.placeholder(tf.float32)

# 把数据转换成 适合 nn 输入的数据格式 ，
# -1代表先不考虑输入的图片例子多少个维度（图片的个数），图片的大小28*28，1：通道
x_image = tf.reshape(xs, [-1, 28, 28, 1])

# patch（卷积映射在图像的范围） 5x5 ,channel is 1 ， output 32 featuremap
#一个卷积计算的结果1维，32个卷积
W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])

#非线性化
# structure is 28x28x32
h_cov1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
# structure is 14x14x32
h_pool1 = max_poo_2x2(h_cov1)


W_conv2 = weight_variable([5, 5, 32, 64])
b_conv2 = bias_variable([64])

# structure is 14x14x64
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
# structure is 7x7x64
h_pool2 = max_poo_2x2(h_conv2)

# full connection
h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])

W_fc1 = weight_variable([7*7*64, 1024])
b_fc1 = bias_variable([1024])

h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

# add dropout
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

# last layer
W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])

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

# define loss
cross_entropy = tf.reduce_mean( - tf.reduce_sum(ys * tf.log(prediction), reduction_indices = [1]) )

train_step = tf.train.AdamOptimizer( 1e-4 ).minimize( cross_entropy )

sess = tf.Session()
sess.run(tf.global_variables_initializer())

for i in range(1000):
    batch_xs, batch_ys = mnist.train.next_batch(100)
    sess.run(train_step, feed_dict = {xs : batch_xs, ys : batch_ys, keep_prob : 0.5})
    if i % 50 == 0:
        print(compute_accuracy(mnist.test.images[:1000], mnist.test.labels[:1000]))

总结：

为什么使用池化：
为得到较小的图，采用步长为2的卷积提取，但信息丢失严重，可以采用步长为1的卷积提取，再池化，得到相同的结果

总体流程：
卷积->池化->卷积->池化->全连接

1. 卷积
   • 一个图片维数[28,28,1]
   • 卷积权重使用正态分布随机生成5*5，颜色通道：1，32个卷积，生成维数为[5, 5, 1, 32]
   • 偏置项为[32]
   • tf.nn.conv2d(x, W, strides = [1, 1, 1, 1], padding = ‘SAME’)
   • 对卷积的结果非线性化：tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)得到28x28x32
2. 池化
   • tf.nn.max_pool(x, ksize = [1, 2, 2, 1], strides = [1, 2, 2, 1], padding = ‘SAME’)
   • 窗口的大小ksize（2*2）
   • 移动步长strides（2*2）
   • padding：SAME，池化可越界，补充0,VALID,不可越界
   • 得到14x14x32
3. 卷积
   • 卷积权重[5, 5, 32, 64]
   • 偏置项为[32]
   • 对卷积的结果非线性化：14x14x64
4. 池化
   • 7x7x64
5. 全连接层
   • [7,7,64]转化[7x7x64]
   • 三层神经网络，使用dropout防止过拟合