背景简介
TensorFlow实现讲解
设计新思路:
1.使用类来记录整个网络:
使用_init_()属性来记录 网络超参数 & 网络框架 & 训练过程
使用一个隐式方法初始化网络参数
2.使用字典存储初始化的各个参数(w&b)
参数初始化新思路:
主程序:
图结构实际实现
Version1:
导入包:
import numpy as np import sklearn.preprocessing as prep import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data import os
关闭日志警告:
级别2是警告信息忽略,级别3是错误信息忽略
# 关闭tensorflow的警告信息 os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
均匀分布参数生成函数:
# 标准均匀分布 def xavier_init(fan_in,fan_out,constant = 1): low = -constant * np.sqrt(6.0/(fan_in+fan_out)) hight = constant * np.sqrt(6.0/(fan_in+fan_out)) return tf.random_uniform((fan_in,fan_out),maxval=hight,minval=low,dtype=tf.float32)
网络类:
class AdditiveGuassianNoiseAutoencoder(): def __init__(self,n_input,n_hidden,transfer_function=tf.nn.softplus, optimzer=tf.train.AdamOptimizer(),scale=0.1): # 网络参数 self.n_input = n_input self.n_hidden = n_hidden self.transfer = transfer_function # 激活函数 self.training_scale = scale # 噪声水平 network_weights = self._initialize_weights() self.weights = network_weights # 网络结构 self.x = tf.placeholder(tf.float32, [None, self.n_input]) self.hidden = self.transfer( tf.add(tf.matmul(self.x + scale * tf.random_normal((n_input,)), self.weights['w1']), self.weights['b1'])) self.reconstruction = tf.add(tf.matmul(self.hidden, self.weights['w2']), self.weights['b2']) # 训练部分 self.cost = 0.5 * tf.reduce_sum(tf.pow(tf.subtract(self.reconstruction, self.x), 2)) self.optimizer = optimzer.minimize(self.cost) init = tf.global_variables_initializer() self.sess = tf.Session() self.sess.run(init) print('begin to run session... ...') def _initialize_weights(self): all_weights = dict() all_weights['w1'] = tf.Variable(xavier_init(self.n_input, self.n_hidden)) all_weights['b1'] = tf.Variable(tf.zeros([self.n_hidden], dtype=tf.float32)) all_weights['w2'] = tf.Variable(xavier_init(self.n_hidden, self.n_input)) all_weights['b2'] = tf.Variable(tf.zeros([self.n_input], dtype=tf.float32)) return all_weights
主函数:
AGN_AC = AdditiveGuassianNoiseAutoencoder(n_input=784, n_hidden=200, transfer_function=tf.nn.softplus, optimzer=tf.train.AdamOptimizer(learning_rate=0.01), scale=0.01) writer = tf.summary.FileWriter(logdir='logs', graph=AGN_AC.sess.graph) writer.close()
图:
VersionV2
添加了命名空间,使节点更为清晰,但实际图结构显得凌乱,由于W&b的节点没有被划归到层节点下的关系
import numpy as np import sklearn.preprocessing as prep import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data import os # 关闭tensorflow的警告信息 os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' # 标准均匀分布 def xavier_init(fan_in,fan_out,constant = 1): low = -constant * np.sqrt(6.0/(fan_in+fan_out)) hight = constant * np.sqrt(6.0/(fan_in+fan_out)) return tf.random_uniform((fan_in,fan_out),maxval=hight,minval=low,dtype=tf.float32) class AdditiveGuassianNoiseAutoencoder(): def __init__(self,n_input,n_hidden,transfer_function=tf.nn.softplus, optimzer=tf.train.AdamOptimizer(),scale=0.1): # 网络参数 self.n_input = n_input self.n_hidden = n_hidden self.transfer = transfer_function # 激活函数 self.training_scale = scale # 噪声水平 network_weights = self._initialize_weights() self.weights = network_weights # 网络结构 with tf.name_scope('Rawinput'): #<--- self.x = tf.placeholder(tf.float32, [None, self.n_input]) with tf.name_scope('NoiseAdder'): #<--- self.scale = tf.placeholder(dtype=tf.float32) #<---使用占位符取代了固定的scale,增加了feed量 self.noise_x = self.x + self.scale * tf.random_normal((n_input,)) #<--- with tf.name_scope('Encoder'): #<--- self.hidden = self.transfer( tf.add(tf.matmul(self.noise_x, self.weights['w1']), self.weights['b1'])) with tf.name_scope('Reconstruction'): #<--- self.reconstruction = tf.add( tf.matmul(self.hidden, self.weights['w2']), self.weights['b2']) # 训练部分 with tf.name_scope('Loss'): #<--- self.cost = 0.5 * tf.reduce_sum(tf.pow(tf.subtract(self.reconstruction, self.x), 2)) with tf.name_scope('Train'): self.optimizer = optimzer.minimize(self.cost) init = tf.global_variables_initializer() self.sess = tf.Session() self.sess.run(init) print('begin to run session... ...') def _initialize_weights(self): all_weights = dict() all_weights['w1'] = tf.Variable(xavier_init(self.n_input, self.n_hidden) ,name='weight1') #<--- all_weights['b1'] = tf.Variable(tf.zeros([self.n_hidden], dtype=tf.float32), name='bias1') all_weights['w2'] = tf.Variable(xavier_init(self.n_hidden, self.n_input), name='weight2') all_weights['b2'] = tf.Variable(tf.zeros([self.n_input], dtype=tf.float32), name='bias2') return all_weights AGN_AC = AdditiveGuassianNoiseAutoencoder(n_input=784, n_hidden=200, transfer_function=tf.nn.softplus, optimzer=tf.train.AdamOptimizer(learning_rate=0.01), scale=0.01) writer = tf.summary.FileWriter(logdir='logs', graph=AGN_AC.sess.graph) writer.close()
图:
Version3
保留字典结构存储W&b
但是把字典key&value生成拆开放在了每一层中
import numpy as np import sklearn.preprocessing as prep import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data import os # 关闭tensorflow的警告信息 os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' # 标准均匀分布 def xavier_init(fan_in,fan_out,constant = 1): low = -constant * np.sqrt(6.0/(fan_in+fan_out)) hight = constant * np.sqrt(6.0/(fan_in+fan_out)) return tf.random_uniform((fan_in,fan_out),maxval=hight,minval=low,dtype=tf.float32) class AdditiveGuassianNoiseAutoencoder(): def __init__(self,n_input,n_hidden,transfer_function=tf.nn.softplus, optimzer=tf.train.AdamOptimizer(),scale=0.1): # 网络参数 self.n_input = n_input self.n_hidden = n_hidden self.transfer = transfer_function # 激活函数 self.training_scale = scale # 噪声水平 self.weights = dict() #<--- # network_weights = self._initialize_weights() # self.weights = network_weights # 网络结构 with tf.name_scope('Rawinput'): self.x = tf.placeholder(tf.float32, [None, self.n_input]) with tf.name_scope('NoiseAdder'): self.scale = tf.placeholder(dtype=tf.float32) self.noise_x = self.x + self.scale * tf.random_normal((n_input,)) with tf.name_scope('Encoder'): self.weights['w1'] = tf.Variable(xavier_init(self.n_input, self.n_hidden), name='weight1') #<--- self.weights['b1'] = tf.Variable(tf.zeros([self.n_hidden], dtype=tf.float32), name='bias1') #<--- self.hidden = self.transfer( tf.add(tf.matmul(self.noise_x, self.weights['w1']), self.weights['b1'])) with tf.name_scope('Reconstruction'): self.weights['w2'] = tf.Variable(xavier_init(self.n_hidden, self.n_input), name='weight2') #<--- self.weights['b2'] = tf.Variable(tf.zeros([self.n_input], dtype=tf.float32), name='bias2') #<--- self.reconstruction = tf.add( tf.matmul(self.hidden, self.weights['w2']), self.weights['b2']) # 训练部分 with tf.name_scope('Loss'): self.cost = 0.5 * tf.reduce_sum(tf.pow(tf.subtract(self.reconstruction, self.x), 2)) with tf.name_scope('Train'): self.optimizer = optimzer.minimize(self.cost) init = tf.global_variables_initializer() self.sess = tf.Session() self.sess.run(init) print('begin to run session... ...') # def _initialize_weights(self): # all_weights = dict() # all_weights['w1'] = tf.Variable(xavier_init(self.n_input, self.n_hidden) ,name='weight1') # all_weights['b1'] = tf.Variable(tf.zeros([self.n_hidden], dtype=tf.float32), name='bias1') # all_weights['w2'] = tf.Variable(xavier_init(self.n_hidden, self.n_input), name='weight2') # all_weights['b2'] = tf.Variable(tf.zeros([self.n_input], dtype=tf.float32), name='bias2') # return all_weights AGN_AC = AdditiveGuassianNoiseAutoencoder(n_input=784, n_hidden=200, transfer_function=tf.nn.softplus, optimzer=tf.train.AdamOptimizer(learning_rate=0.01), scale=0.01) writer = tf.summary.FileWriter(logdir='logs', graph=AGN_AC.sess.graph) writer.close()
图:
总结:
1.参数变量使用字典保存提升规整性;
2.参数变量生成仍然要放在层中,可视化效果更好。
降噪自编码器完整程序
import numpy as np import sklearn.preprocessing as prep import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data import os # 关闭tensorflow的警告信息 os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' # 标准均匀分布 def xavier_init(fan_in,fan_out,constant = 1): low = -constant * np.sqrt(6.0/(fan_in+fan_out)) hight = constant * np.sqrt(6.0/(fan_in+fan_out)) return tf.random_uniform((fan_in,fan_out),maxval=hight,minval=low,dtype=tf.float32) class AdditiveGuassianNoiseAutoencoder(): def __init__(self,n_input,n_hidden,transfer_function=tf.nn.softplus, optimzer=tf.train.AdamOptimizer(),scale=0.1): # 网络参数 self.n_input = n_input self.n_hidden = n_hidden self.transfer = transfer_function # 激活函数 self.training_scale = scale # 噪声水平 self.weights = dict() # network_weights = self._initialize_weights() # self.weights = network_weights # 网络结构 with tf.name_scope('Rawinput'): self.x = tf.placeholder(tf.float32, [None, self.n_input]) with tf.name_scope('NoiseAdder'): self.scale = tf.placeholder(dtype=tf.float32) self.noise_x = self.x + self.scale * tf.random_normal((n_input,)) with tf.name_scope('Encoder'): self.weights['w1'] = tf.Variable(xavier_init(self.n_input, self.n_hidden), name='weight1') # <--- self.weights['b1'] = tf.Variable(tf.zeros([self.n_hidden], dtype=tf.float32), name='bias1') self.hidden = self.transfer( tf.add(tf.matmul(self.noise_x, self.weights['w1']), self.weights['b1'])) with tf.name_scope('Reconstruction'): self.weights['w2'] = tf.Variable(xavier_init(self.n_hidden, self.n_input), name='weight2') # <--- self.weights['b2'] = tf.Variable(tf.zeros([self.n_input], dtype=tf.float32), name='bias2') self.reconstruction = tf.add( tf.matmul(self.hidden, self.weights['w2']), self.weights['b2']) # 训练部分 with tf.name_scope('Loss'): self.cost = 0.5 * tf.reduce_sum(tf.pow(tf.subtract(self.reconstruction, self.x), 2)) with tf.name_scope('Train'): self.optimizer = optimzer.minimize(self.cost) init = tf.global_variables_initializer() self.sess = tf.Session() self.sess.run(init) print('begin to run session... ...') def partial_fit(self, X): ''' 训练并计算cost :param X: :return: ''' cost, opt = self.sess.run([self.cost, self.optimizer], feed_dict={self.x:X, self.scale:self.training_scale}) return cost def calc_cost(self, X): ''' 不训练,只计算cost :param X: :return: ''' return self.sess.run(self.cost, feed_dict={self.x: X, self.scale: self.training_scale}) # 数据集预处理 def standard_scale(X_train, X_test): #<-----数据集预处理部分 ''' 0均值,1标准差 :param X_train: :param X_test: :return: ''' # 根据预估的训练集的参数生成预处理器 preprocessor = prep.StandardScaler().fit(X_train) X_train = preprocessor.transform(X_train) X_test = preprocessor.transform(X_test) return X_train, X_test def get_random_block_from_data(data, batch_size): ''' 随机取一个batch的数据 :param data: :param batch_size: :return: ''' start_index = np.random.randint(0, len(data) - batch_size) return data[start_index:(start_index+batch_size)] # 展示计算图 AGN_AC = AdditiveGuassianNoiseAutoencoder(n_input=784, n_hidden=200, transfer_function=tf.nn.softplus, optimzer=tf.train.AdamOptimizer(learning_rate=0.01), scale=0.01) writer = tf.summary.FileWriter(logdir='logs', graph=AGN_AC.sess.graph) writer.close() # 读取数据 mnist = input_data.read_data_sets('../Mnist_data/', one_hot=True) X_train, X_test = standard_scale(mnist.train.images, mnist.test.images) n_samples = int(mnist.train.num_examples) # 训练样本总数 training_epochs = 20 # 训练轮数,1轮等于n_samples/batch_size batch_size = 128 # batch容量 display_step = 1 # 展示间隔 # 训练 for epoch in range(training_epochs): avg_cost = 0 # 平均损失 total_batch = int(n_samples/batch_size) # 每一轮中step总数 for i in range(total_batch): batch_xs = get_random_block_from_data(X_train, batch_size) cost = AGN_AC.partial_fit(batch_xs) avg_cost += cost / batch_size avg_cost /= total_batch if epoch % display_step == 0: print('epoch : %04d, cost = %.9f' % (epoch+1, avg_cost)) # 计算测试集上的cost print('Total coat:', str(AGN_AC.calc_cost(X_test)))
引入了数据预处理机制:
import sklearn.preprocessing as prep # 数据集预处理 def standard_scale(X_train, X_test): #<-----一个新的尝试 ''' 0均值,1标准差 :param X_train: :param X_test: :return: ''' # 根据预估的训练集的参数生成预处理器 preprocessor = prep.StandardScaler().fit(X_train) X_train = preprocessor.transform(X_train) X_test = preprocessor.transform(X_test) return X_train, X_test
因为就是个范例而已,所以并没有加入更多的步骤。
输出:
Python 3.6.0 |Anaconda 4.3.1 (64-bit)| (default, Dec 23 2016, 12:22:00)
Type "copyright", "credits" or "license" for more information.
IPython 5.1.0 -- An enhanced Interactive Python.
? -> Introduction and overview of IPython's features.
%quickref -> Quick reference.
help -> Python's own help system.
object? -> Details about 'object', use 'object??' for extra details.
PyDev console: using IPython 5.1.0
Running /home/hellcat/PycharmProjects/data_analysis/TensorFlow/autoencodes/denoise.py
Backend Qt5Agg is interactive backend. Turning interactive mode on.
begin to run session... ...
Extracting ../Mnist_data/train-images-idx3-ubyte.gz
Extracting ../Mnist_data/train-labels-idx1-ubyte.gz
Extracting ../Mnist_data/t10k-images-idx3-ubyte.gz
Extracting ../Mnist_data/t10k-labels-idx1-ubyte.gz
epoch : 0001, cost = 172728.840323210
epoch : 0002, cost = 384090.340043217
epoch : 0003, cost = 1424137.733514817
epoch : 0004, cost = 252195.476644165
epoch : 0005, cost = 1989602.406287275
epoch : 0006, cost = 82078.567135613
epoch : 0007, cost = 4571607.288953234
epoch : 0008, cost = 12936386.999440582
epoch : 0009, cost = 192551.124642752
epoch : 0010, cost = 40848.185927740
epoch : 0011, cost = 2998.114711095
epoch : 0012, cost = 15210.583374379
epoch : 0013, cost = 38411.792979990
epoch : 0014, cost = 5556.733809144
epoch : 0015, cost = 35625.806443790
epoch : 0016, cost = 1274942.135287910
epoch : 0017, cost = 214436.171889868
epoch : 0018, cost = 29740.634501637
epoch : 0019, cost = 1136.356513888
epoch : 0020, cost = 2248.695473340
Total coat: 4.2886e+06