自动编码器是一种比较好理解的神经网络结构。它是一种无监督的学习特征方法(从自己到自己)。下面以一个例子来介绍。
使用的数据集:[数据集](https://archive.ics.uci.edu/ml/machine-learning-databases/00310/),下载后进行解压。
首先给出完整代码:
#coding=utf-8
#用sklearn对数据集进行处理
import tensorflow as tf
from sklearn.preprocessing import scale#
import pandas as pd
import numpy as np
training_data = pd.read_csv("UJIndoorLoc/trainingData.csv",header=0)
training_x = scale(np.asarray(training_data.ix[:,0:520]))
training_y = np.asarray(training_data["BUILDINGID"].map(str) + training_data["FLOOR"].map(str))
training_y = np.asarray(pd.get_dummies(training_y))
test_dataset = pd.read_csv("UJIndoorLoc/validationData.csv",header=0)
test_x = scale(np.asarray(test_dataset.ix[:,0:520]))
test_y = np.asarray(test_dataset["BUILDINGID"].map(str) + test_dataset["FLOOR"].map(str))
test_y = np.asarray(pd.get_dummies(test_y))
output = training_y.shape[1]
X = tf.placeholder(tf.float32,shape=[None,520])#网络输入
Y = tf.placeholder(tf.float32,[None,output])
#定义神经网络
def neural_network():
#Encoder
e_w_1 = tf.Variable(tf.truncated_normal([520,256],stddev=0.1))
e_b_1 = tf.Variable(tf.constant(0.0,shape=[256]))
e_w_2 = tf.Variable(tf.truncated_normal([256,128],stddev=0.1))
e_b_2 = tf.Variable(tf.constant(0.0,shape=[128]))
e_w_3 = tf.Variable(tf.truncated_normal([128,64],stddev=0.1))
e_b_3 = tf.Variable(tf.constant(0.0,shape=[64]))
#Decoder
d_w_1 = tf.Variable(tf.truncated_normal([64,128],stddev=0.1))
d_b_1 = tf.Variable(tf.constant(0.0,shape=[128]))
d_w_2 = tf.Variable(tf.truncated_normal([128,256],stddev=0.1))
d_b_2 = tf.Variable(tf.constant(0.0,shape=[256]))
d_w_3 = tf.Variable(tf.truncated_normal([256,520],stddev=0.1))
d_b_3 = tf.Variable(tf.constant(0.0,shape=[520]))
#DNN
w_1 = tf.Variable(tf.truncated_normal([64,128],stddev=0.1))
b_1= tf.Variable(tf.constant(0.0,shape=[128]))
w_2= tf.Variable(tf.truncated_normal([128,128],stddev=0.1))
b_2 = tf.Variable(tf.constant(0.0,shape=[128]))
w_3 = tf.Variable(tf.truncated_normal([128,output],stddev=0.1))
b_3 = tf.Variable(tf.constant(0.0,shape=[output]))
#####
layer_1 = tf.nn.tanh(tf.add(tf.matmul(X,e_w_1),e_b_1))
layer_2 = tf.nn.tanh(tf.add(tf.matmul(layer_1,e_w_2),e_b_2))
encoded = tf.nn.tanh(tf.add(tf.matmul(layer_2,e_w_3),e_b_3))
layer_4 = tf.nn.tanh(tf.add(tf.matmul(encoded,d_w_1),d_b_1))
layer_5 = tf.nn.tanh(tf.add(tf.matmul(layer_4,d_w_2),d_b_2))
decoded = tf.nn.tanh(tf.add(tf.matmul(layer_5,d_w_3),d_b_3))
layer_7 = tf.nn.relu(tf.add(tf.matmul(encoded,w_1),b_1))
layer_8 = tf.nn.relu(tf.add(tf.matmul(layer_7,w_2),b_2))
out = tf.nn.softmax(tf.add(tf.matmul(layer_8,w_3),b_3))
return (decoded, out)
#训练神经网络
def train_neural_networks():
decoded, predict_output = neural_network()
us_cost_function = tf.reduce_mean(tf.pow(X-decoded,2))
s_cost_function = -tf.reduce_sum(Y * tf.log(predict_output) )
us_optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01).minimize(us_cost_function)
s_optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01).minimize(s_cost_function)
correct_prediction = tf.equal(tf.argmax(predict_output, 1), tf.argmax(Y,1))
accuray = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
training_epochs = 20
batch_size = 10
total_batches = training_data.shape[0]
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
#autoencoder一种非监督学习算法
for epoch in range(training_epochs):
epoch_costs = np.empty(0)
for b in range(total_batches):
offset = (b* batch_size)%(training_x.shape[0] - batch_size)
batch_x = training_x[offset:(offset+batch_size),:]
_,c = sess.run([us_optimizer,us_cost_function],feed_dict={X:batch_x})
epoch_costs = np.append(epoch_costs,c)
print("Eopch:",epoch," Loss: ",np.mean(epoch_costs))
print("----------------------------------------------")
# ---------------- Training NN - Supervised Learning ------------------ #
for epoch in range(training_epochs):
epoch_costs = np.empty(0)
for b in range(total_batches):
offset = (b * batch_size) % (training_x.shape[0]- batch_size)
batch_x = training_x[offset:(offset+batch_size), :]
batch_y = training_y[offset:(offset+batch_size), :]
_,c = sess.run([s_optimizer,s_cost_function],feed_dict={X:batch_x,Y:batch_y})
epoch_costs = np.append(epoch_costs,c)
accuray_in_train_set = sess.run(accuray, feed_dict={X:training_x,Y:training_y})
accuray_in_test_set = sess.run(accuray,feed_dict={X: test_x,Y: test_y})
print("Epoch: ",epoch, " Loss:",np.mean(epoch_costs)," Accuracy: ", accuray_in_train_set, ' ', accuray_in_test_set)
train_neural_networks()
(1)加载数据集;
training_data = pd.read_csv("UJIndoorLoc/trainingData.csv",header=0)
training_x = scale(np.asarray(training_data.ix[:,0:520]))
training_y = np.asarray(training_data["BUILDINGID"].map(str) + training_data["FLOOR"].map(str))
training_y = np.asarray(pd.get_dummies(training_y))
test_dataset = pd.read_csv("UJIndoorLoc/validationData.csv",header=0)
test_x = scale(np.asarray(test_dataset.ix[:,0:520]))
test_y = np.asarray(test_dataset["BUILDINGID"].map(str) + test_dataset["FLOOR"].map(str))
test_y = np.asarray(pd.get_dummies(test_y))
这里,我们使用了sklearn对数据集进行处理。接下来定义神经网络:
(2)定义神经网络
def neural_network():
#Encoder(3层)
e_w_1 = tf.Variable(tf.truncated_normal([520,256],stddev=0.1))
e_b_1 = tf.Variable(tf.constant(0.0,shape=[256]))
e_w_2 = tf.Variable(tf.truncated_normal([256,128],stddev=0.1))
e_b_2 = tf.Variable(tf.constant(0.0,shape=[128]))
e_w_3 = tf.Variable(tf.truncated_normal([128,64],stddev=0.1))
e_b_3 = tf.Variable(tf.constant(0.0,shape=[64]))
#Decoder(3层)
d_w_1 = tf.Variable(tf.truncated_normal([64,128],stddev=0.1))
d_b_1 = tf.Variable(tf.constant(0.0,shape=[128]))
d_w_2 = tf.Variable(tf.truncated_normal([128,256],stddev=0.1))
d_b_2 = tf.Variable(tf.constant(0.0,shape=[256]))
d_w_3 = tf.Variable(tf.truncated_normal([256,520],stddev=0.1))
d_b_3 = tf.Variable(tf.constant(0.0,shape=[520]))
#DNN(3层)
w_1 = tf.Variable(tf.truncated_normal([64,128],stddev=0.1))
b_1= tf.Variable(tf.constant(0.0,shape=[128]))
w_2= tf.Variable(tf.truncated_normal([128,128],stddev=0.1))
b_2 = tf.Variable(tf.constant(0.0,shape=[128]))
w_3 = tf.Variable(tf.truncated_normal([128,output],stddev=0.1))
b_3 = tf.Variable(tf.constant(0.0,shape=[output]))
#####
layer_1 = tf.nn.tanh(tf.add(tf.matmul(X,e_w_1),e_b_1))
layer_2 = tf.nn.tanh(tf.add(tf.matmul(layer_1,e_w_2),e_b_2))
encoded = tf.nn.tanh(tf.add(tf.matmul(layer_2,e_w_3),e_b_3))
layer_4 = tf.nn.tanh(tf.add(tf.matmul(encoded,d_w_1),d_b_1))
layer_5 = tf.nn.tanh(tf.add(tf.matmul(layer_4,d_w_2),d_b_2))
decoded = tf.nn.tanh(tf.add(tf.matmul(layer_5,d_w_3),d_b_3))
layer_7 = tf.nn.relu(tf.add(tf.matmul(encoded,w_1),b_1))
layer_8 = tf.nn.relu(tf.add(tf.matmul(layer_7,w_2),b_2))
out = tf.nn.softmax(tf.add(tf.matmul(layer_8,w_3),b_3))
return (decoded, out)
三层编码,三层解码,两层全连接,一个输出层;输出为解码器的输出与与猜测输出。
训练过程中先进行自编码的训练,用于特征提取,然后用于训练、测试网络的精确率。
运行结果如下: