在MNIST数据集上训练一个卷积网络自编码器

版权声明：转载请表明来源，多谢！ https://blog.csdn.net/kingfoulin/article/details/81315937

  首先你肯定知道了什么是自编码，一般我们常见的自编码是使用的多层感知机来实现的，也就是多层的全连接神经网络结构。本小记中我们使用CNN实现一个七层的卷积神经网络构成的自编码器。
   自编码器使用很广泛，我觉得它的思想就是同一空间的数据操作的最优结果，数据经过压缩（数据向前传播到达中间的层所得到的结果），然后数据的解压缩过程(数据到达网络的输出)，回到原来的空间。这个思想很关键，利用这个思想我们可以对元数据进行一些去噪操作，等等。

   训练程序：

#!user/bin/python
# _*_ coding:utf-8 _*_

import tensorflow as tf
import matplotlib.pyplot as pt
import numpy as np
from tensorflow.examples.tutorials.mnist import input_data

mnist = input_data.read_data_sets('../MNIST_data/', one_hot=True)
batch_size = 64
sess = tf.InteractiveSession()

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

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

def conv2d(x, w):
return tf.nn.conv2d(x, w, strides=[1, 2, 2, 1], padding='SAME')

# 四层结构的反卷积操作

inputdata = tf.placeholder(tf.float32, shape=[None, 28 * 28])
real_label = tf.placeholder(tf.float32, shape=[None, 28 * 28])  # 正确的标签

input_image = tf.reshape(inputdata, [-1, 28, 28, 1])

# layer1
w_conv1 = weight_varibales([5, 5, 1, 32])
b_conv1 = bias_variables([32])
h_conv1 = tf.nn.relu(conv2d(input_image, w_conv1) + b_conv1)
# 14*14*32

# layer2
w_conv2 = weight_varibales([5, 5, 32, 64])
b_conv2 = bias_variables([64])
h_conv2 = tf.nn.relu(conv2d(h_conv1, w_conv2) + b_conv2)
# 7*7*64

# layer3
w_conv3 = weight_varibales([3, 3, 64, 128])
b_conv3 = bias_variables([128])
h_conv3 = tf.nn.relu(conv2d(h_conv2, w_conv3) + b_conv3)
# 4*4*128

# decode_layer 1 input 4*4*128
w_decode1 = weight_varibales([3, 3, 64, 128])
h_decode1 = tf.nn.conv2d_transpose(h_conv3, w_decode1, [batch_size, 7, 7, 64], [1, 2, 2, 1], padding="SAME")
h_out1 = tf.nn.relu(h_decode1)

# decode layer 2 input 7*7*64
w_decode2 = weight_varibales([3, 3, 32, 64])
h_decode2 = tf.nn.conv2d_transpose(h_out1, w_decode2, [batch_size, 14, 14, 32], [1, 2, 2, 1], padding='SAME')
h_out2 = tf.nn.relu(h_decode2)

# decode layer 3 input 14*14*32
w_decode3 = weight_varibales([5, 5, 1, 32])
h_decode3 = tf.nn.conv2d_transpose(h_out2, w_decode3, [batch_size, 28, 28, 1], [1, 2, 2, 1], padding='SAME')
h_out = tf.nn.relu(h_decode3)
# the network's output is 28*28*1

output = tf.reshape(h_out, [-1, 28 * 28])

loss = tf.reduce_mean(pow(output - inputdata, 2))
train_step = tf.train.AdamOptimizer(0.001).minimize(loss)
tf.summary.scalar('loss', loss)

saver = tf.train.Saver()
write = tf.summary.FileWriter('pic/',sess.graph) # 注意这里的后面是需要加上sess.graph的，不然的话无法在tensorflow显示流程图
merge = tf.summary.merge_all()

sess.run(tf.initialize_all_variables())

for i in range(5000):
images, labels = mnist.train.next_batch(batch_size)
result = sess.run(merge, {inputdata: images})
write.add_summary(result)

if i % 100 == 0:
print('loss: ', loss.eval({inputdata:images}))

sess.run(train_step, {inputdata: images})

saver.save(sess, 'model/')

   检测程序：

#!user/bin/python
# _*_ coding:utf-8 _*_

import tensorflow as tf
import matplotlib.pyplot as pt
import numpy as np
from tensorflow.examples.tutorials.mnist import input_data

mnist = input_data.read_data_sets('../MNIST_data/', one_hot=True)
batch_size = 2
sess = tf.InteractiveSession()

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

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

def conv2d(x, w):
return tf.nn.conv2d(x, w, strides=[1, 2, 2, 1], padding='SAME')

# 四层结构的反卷积操作

inputdata = tf.placeholder(tf.float32, shape=[None, 28 * 28])
real_label = tf.placeholder(tf.float32, shape=[None, 28 * 28])  # 正确的标签

input_image = tf.reshape(inputdata, [-1, 28, 28, 1])

# layer1
w_conv1 = weight_varibales([5, 5, 1, 32])
b_conv1 = bias_variables([32])
h_conv1 = tf.nn.relu(conv2d(input_image, w_conv1) + b_conv1)
# 14*14*32

# layer2
w_conv2 = weight_varibales([5, 5, 32, 64])
b_conv2 = bias_variables([64])
h_conv2 = tf.nn.relu(conv2d(h_conv1, w_conv2) + b_conv2)
# 7*7*64

# layer3
w_conv3 = weight_varibales([3, 3, 64, 128])
b_conv3 = bias_variables([128])
h_conv3 = tf.nn.relu(conv2d(h_conv2, w_conv3) + b_conv3)
# 4*4*128

# decode_layer 1 input 4*4*128
w_decode1 = weight_varibales([3, 3, 64, 128])
h_decode1 = tf.nn.conv2d_transpose(h_conv3, w_decode1, [batch_size, 7, 7, 64], [1, 2, 2, 1], padding="SAME")
h_out1 = tf.nn.relu(h_decode1)

# decode layer 2 input 7*7*64
w_decode2 = weight_varibales([3, 3, 32, 64])
h_decode2 = tf.nn.conv2d_transpose(h_out1, w_decode2, [batch_size, 14, 14, 32], [1, 2, 2, 1], padding='SAME')
h_out2 = tf.nn.relu(h_decode2)

# decode layer 3 input 14*14*32
w_decode3 = weight_varibales([5, 5, 1, 32])
h_decode3 = tf.nn.conv2d_transpose(h_out2, w_decode3, [batch_size, 28, 28, 1], [1, 2, 2, 1], padding='SAME')
h_out = tf.nn.relu(h_decode3)
# the network's output is 28*28*1

output = tf.reshape(h_out, [-1, 28 * 28])

loss = tf.reduce_mean(pow(output - inputdata, 2))
train_step = tf.train.AdamOptimizer(0.001).minimize(loss)

saver = tf.train.Saver()

saver.restore(sess, 'model/')

images, labels = mnist.test.next_batch(batch_size)
print(output.shape)
pt.subplot(1, 2, 1)
pt.imshow(images[0].reshape(28, 28), cmap='gray')
pt.subplot(1, 2, 2)
pt.imshow(output.eval({inputdata: images})[0].reshape(28, 28), cmap='gray')
pt.show()

由上面的程序中，我们在测试的时候，使用的

batch_size=2

，主要测试时候可视化数据基本要求不多。因此我们就使用batch_size很小的数据集进行测试。

阅读更多