自編碼器簡介
代碼及詳細(xì)注釋
#!/usr/bin/env python2
# -*- coding: utf-8 -*-
"""
Created on Fri Mar 31 16:05:38 2017
@author: mml
"""
import numpy as np
# 數(shù)據(jù)預(yù)處理模塊
import sklearn.preprocessing as prep
import tensorflow as tf
# 使用MNIST數(shù)據(jù)集
from tensorflow.examples.tutorials.mnist import input_data
# 參數(shù)初始化方法
# 自動根據(jù)某一層網(wǎng)絡(luò)的輸入模闲，輸出節(jié)點數(shù)量自動調(diào)整最合適分布
# fan_in輸入節(jié)點數(shù)量 fan_out輸出節(jié)點數(shù)量
def xavier_init(fan_in,fan_out,constant = 1):
low = -constant*np.sqrt(6.0/(fan_in+fan_out))
high = constant*np.sqrt(6.0/(fan_in+fan_out))
# tf.random_uniform創(chuàng)建一個low到high之間的均勻分布
return tf.random_uniform((fan_in,fan_out),minval = low,maxval = high,dtype = tf.float32)
class AdditiveGaussianNoiseAutoencoder(object):
# 構(gòu)建函數(shù) 輸入變量數(shù)暂筝，隱層節(jié)點數(shù) 激活函數(shù) 優(yōu)化器 scale高斯噪聲系數(shù)
def __init__(self,n_input,n_hidden,transfer_function=tf.nn.softplus,
optimizer=tf.train.AdamOptimizer(),scale=0.1):
# 輸入變量數(shù)
self.n_input = n_input
# 隱層節(jié)點數(shù)
self.n_hidden = n_hidden
# 激活函數(shù)
self.transfer = transfer_function
self.scale = tf.placeholder(tf.float32)
self.training_scale = scale
# 參數(shù)初始化方法
network_weights = self._initialize_weights()
self.weights = network_weights
# 為輸入x創(chuàng)建一個維度為n_input的placeholder
self.x = tf.placeholder(tf.float32,[None,self.n_input])
# 隱含層提取特征過程
# scale*tf.random_normal((n_input,)）產(chǎn)生高斯噪聲
# self.x + scale*tf.random_normal((n_input,)) 為輸入加上高斯噪聲
# tf.matmul(self.x + scale*tf.random_normal((n_input,)),self.weights['w1']) 加入噪聲后的輸入乘以權(quán)重
# tf.add(tf.matmul(self.x + scale*tf.random_normal((n_input,)),self.weights['w1']),self.weights['b1'])) 最后加上偏置
# self.transfer() 對結(jié)果進(jìn)行激活函數(shù)處理
self.hidden = self.transfer(tf.add(tf.matmul(self.x + scale*tf.random_normal
((n_input,)),self.weights['w1']),
self.weights['b1']))
# 經(jīng)過隱含層提取特征后游沿，我們需要在輸出層進(jìn)行數(shù)據(jù)復(fù)原重建操作
# 重構(gòu)層直接把隱含層輸出乘以輸出層權(quán)重并加上偏置即可
self.reconstruction = tf.add(tf.matmul(self.hidden,
self.weights['w2']),self.weights['b2'])
# 自編碼器的損失函數(shù) 平方誤差作為cost
# tf.subtract(self.reconstruction,self.x) 重構(gòu)后的輸出和輸入相減
# tf.pow求差的平方
# tf.reduce_sum求所有平方誤差和
self.cost = 0.5 * tf.reduce_sum(tf.pow(tf.subtract(
self.reconstruction,self.x),2.0))
# 定義優(yōu)化方法左电，對cost進(jìn)行優(yōu)化
self.optimizer = optimizer.minimize(self.cost)
# 全局參數(shù)初始化
init = tf.global_variables_initializer()
# 創(chuàng)建Session
self.sess = tf.Session()
self.sess.run(init)
# 參數(shù)初始化函數(shù)
def _initialize_weights(self):
# 創(chuàng)建一個所有參數(shù)的字典
all_weights = dict()
# w1使用前面的xavier_init初始化
all_weights['w1'] = tf.Variable(xavier_init(self.n_input,self.n_hidden))
# 其它都初始化為0
all_weights['b1'] = tf.Variable(tf.zeros([self.n_hidden],dtype = tf.float32))
all_weights['w2'] = tf.Variable(tf.zeros([self.n_hidden,self.n_input],dtype = tf.float32))
all_weights['b2'] = tf.Variable(tf.zeros([self.n_input],dtype = tf.float32))
return all_weights
# 定義一個batch數(shù)據(jù)進(jìn)行訓(xùn)練并返回當(dāng)前cost
def partial_fit(self,X):
# 讓Session執(zhí)行計算流圖節(jié)點cost和optimizer
# feed_dict為輸入數(shù)據(jù)X和噪聲系數(shù)
cost,opt = self.sess.run((self.cost,self.optimizer),
feed_dict = {self.x:X,self.scale:self.training_scale})
return cost
# 還需要一個只計算cost不訓(xùn)練的函數(shù)
def calc_total_cost(self,X):
# 讓Session只觸發(fā)計算流圖節(jié)點self.cost
return self.sess.run(self.cost,feed_dict = {self.x:X,self.scale:self.training_scale})
# 還需函數(shù)返回隱含層輸出結(jié)果（即提取的特征）
def transform(self,X):
# Session觸發(fā)計算節(jié)點hidden
return self.sess.run(self.hidden,feed_dict = {self.x:X,
self.scale:self.training_scale})
# 定義函數(shù)進(jìn)行單獨重建（輸入為隱含層輸出）
def generate(self,hidden = None):
if hidden is None:
hidden = np.random.normal(size = self.weights['b1'])
return self.sess.run(self.reconstruction,feed_dict = {self.hidden:hidden})
# 定義完整的重建穴亏，包括前面的transform和reconstruction
def reconstruct(self,X):
return self.sess.run(self.reconstruction,feed_dict = {self.x:X,self.scale:self.training_scale})
# 獲取隱含層參數(shù)
def getWeights(self):
return self.sess.run(self.weights['w1'])
def getBiases(self):
return self.sess.run(self.weights['b1'])
# 載入MINIST數(shù)據(jù)集
mnist = input_data.read_data_sets('MNIST_data',one_hot = True)
# 對訓(xùn)練和測試數(shù)據(jù)進(jìn)行標(biāo)準(zhǔn)化處理
# 標(biāo)準(zhǔn)化讓數(shù)據(jù)變成0均值妄迁，且標(biāo)準(zhǔn)差為1的分布
def standard_scale(X_train,X_test):
preprocessor = prep.StandardScaler().fit(X_train)
X_train = preprocessor.transform(X_train)
X_test = preprocessor.transform(X_test)
return X_train,X_test
# 定義獲取隨機(jī)block數(shù)據(jù)的方法
def get_random_block_from_data(data,batch_size):
# 取一個隨機(jī)整數(shù)
start_index = np.random.randint(0,len(data)-batch_size)
# 順序取到一個batch_size的數(shù)據(jù)
return data[start_index:(start_index + batch_size)]
# 使用之前的標(biāo)準(zhǔn)化函數(shù)對訓(xùn)練集和測試集進(jìn)行標(biāo)準(zhǔn)化處理
X_train,X_test = standard_scale(mnist.train.images,mnist.test.images)
# 總訓(xùn)練樣本數(shù)
n_samples = int(mnist.train.num_examples)
# 最大訓(xùn)練輪數(shù)
training_epochs = 20
# batchsize
batch_size = 128
# 每一輪顯示一次cost
display_step = 1
# 創(chuàng)建AGN實例
# 輸入784（mnist數(shù)據(jù)28*28）
autoencoder = AdditiveGaussianNoiseAutoencoder(n_input = 784,
n_hidden = 200,
transfer_function = tf.nn.softplus,
optimizer = tf.train.AdagradOptimizer(learning_rate = 0.001),scale = 0.01)
# 開始真正的訓(xùn)練過程
for epoch in range(training_epochs):
avg_cost = 0.
# 計算總共的batch數(shù)
total_batch = int(n_samples / batch_size)
for i in range(total_batch):
# 使用get_random_block_from_data獲取隨機(jī)batch數(shù)據(jù)
batch_xs = get_random_block_from_data(X_train,batch_size)
# 使用partial_fit進(jìn)行訓(xùn)練卵渴，并返回cost
cost = autoencoder.partial_fit(batch_xs)
avg_cost += cost / n_samples * batch_size
if epoch % display_step == 0:
print "Epoch:",'%04d' % (epoch + 1),"cost=","{:.9f}".format(avg_cost)
print "Total cost: " + str(autoencoder.calc_total_cost(X_test))
結(jié)果
