【神经网络】AutoEncoder自编码器
1.基本功能
自编码器最大的特点是输入与输出的节点数是一致的，其功能并不是进行分类或者是回归预测，而是将原始数据转换为一种特殊的特征表示方式，在训练的过程中，依然采用反向传播算法进行参数优化，目标函数为使输出与输入之间的误差最小化。
基本功能如图，如图所示，普通神经网络是训练类别（或者离散值）与输入之间的映射关系，而AutoEncoder是将input通过encoder进行编码得到code，然后再使用decoder进行解码，得到原始数据相同维度的表示方式reconstruction，进而通过反向传播优化input与rexonstruction之间的误差。
到目前为止，只是获取到了一个新的特征向量，因此需要连接其他网络结构实现预测分类等功能。通过训练，这个特征可以在最大程度上表示原始数据。最终我们的模型是AutoEncoder+分类器。如图分别为两种监督训练方式：
第一张图片表示在训练prediction与label过程中对整个网络进行参数调节；第二张图片表示在训练过程中仅调节分类器，不修改编码器内部参数。第一种属于End-To-End算法，效果较好。
2.算法实现
（1）导入包
import tflearn.datasets.mnist as mnist
import tflearn
import numpy as np
import matplotlib.pyplot as plt
from tflearn.layers.core import fully_connected, input_data（2）构建Encoder和Decoder
X, Y, testX, testY = mnist.load_data(one_hot=True)
encoder = input_data(shape=[None, 784])
encoder = fully_connected(encoder, 256)
encoder = fully_connected(encoder, 64)
decoder = fully_connected(encoder, 256)
decoder = fully_connected(encoder, 784)
net = tflearn.regression(decoder, optimizer='adam', loss='mean_square',
learning_rate=0.001, metric=None)
model = tflearn.DNN(net)
model.fit(X, X,n_epoch=3, validation_set=(testX, testX),show_metric=True,batch_size=256)（3）测试Encoder结果
print('\nTest encoding of X[0]: ')
encoding_model = tflearn.DNN(encoder, session=model.session)
print(encoding_model.predict([X[0]]))通过调用model.fit()的会话实现权重共享。
（4）可视化结果
print('\nVisualizing results after being encoder and decoder')
testX = tflearn.data_utils.shuffle(testX)[0]
encoder_decoder = model.predict(testX)
# plt
f, a = plt.subplots(2, 10, figsize=(10, 2))
for i in range(10):
temp = [[ii, ii, ii] for ii in list(testX[i])]
a[0][i].imshow(np.reshape(temp, (28, 28, 3)))
temp = [[ii, ii, ii] for ii in list(encoder_decoder[i])]
a[1][i].imshow(np.reshape(temp, (28, 28, 3)))
f.show()
plt.draw()
plt.waitforbuttonpress()（5）连接分类器
此处以Alexnet为例进行测试，效果一般，主要是理解过程
from tflearn.layers.estimator import regression
from tflearn.layers.conv import conv_2d, max_pool_2d
from tflearn.layers.normalization import local_response_normalization
from tflearn.layers.core import dropout
import tensorflow as tf
tf.reset_default_graph()
codeX = model.predict(X)
codeX = codeX.reshape([-1, 28, 28, 1])
network = input_data(shape=[None, 28, 28, 1])
network = conv_2d(network, nb_filter=96, filter_size=8, strides=2, activation='relu')
network = max_pool_2d(network, kernel_size=3, strides=1)
network = local_response_normalization(network)
network = conv_2d(network, nb_filter=256, filter_size=5, strides=1, activation='relu')
network = max_pool_2d(network, kernel_size=3, strides=1)
network = local_response_normalization(network)
network = conv_2d(network, nb_filter=384, filter_size=3, activation='relu')
network = conv_2d(network, nb_filter=384, filter_size=3, activation='relu')
network = conv_2d(network, nb_filter=256, filter_size=3, activation='relu')
network = max_pool_2d(network, kernel_size=3, strides=1)
network = fully_connected(network, 2048, activation='tanh')
network = dropout(network, 0.5)
network = fully_connected(network, 2048, activation='tanh')
network = dropout(network, 0.5)
network = fully_connected(network, 10, activation='softmax')
network = regression(network, optimizer='momentum', loss='categorical_crossentropy', learning_rate=0.001)
model_alexnet = tflearn.DNN(network)
model_alexnet.fit(codeX, Y, n_epoch=100, validation_set=0.2, show_metric=True)
参考链接：http://blog.csdn.net/u010555688/article/details/24438311