Tensorflow 循环神经网络 基本 RNN 和 LSTM 网络 拟合、预测sin曲线
时序预测一直是比较重要的研究问题,在统计学中我们有各种的模型来解决时间序列问题,但是最近几年比较火的深度学习中也有能解决时序预测问题的方法,另外在深度学习领域中时序预测算法可以解决自然语言问题等。
在网上找到了 tensorflow 中 RNN 和 LSTM 算法预测 sin 曲线的代码,效果不错。
LSTM:
#encoding:UTF-8 import random
import numpy as np
import tensorflow as tf
from tensorflow.contrib.rnn.python.ops import core_rnn
from tensorflow.contrib.rnn.python.ops import core_rnn_cell def build_data(n):
xs = []
ys = []
for i in range(0, 2000):
k = random.uniform(1, 50) x = [[np.sin(k + j)] for j in range(0, n)]
y = [np.sin(k + n)] # x[i] = sin(k + i) (i = 0, 1, ..., n-1)
# y[i] = sin(k + n)
xs.append(x)
ys.append(y) train_x = np.array(xs[0: 1500])
train_y = np.array(ys[0: 1500])
test_x = np.array(xs[1500:])
test_y = np.array(ys[1500:])
return (train_x, train_y, test_x, test_y) length = 10
time_step_size = length
vector_size = 1
batch_size = 10
test_size = 10 # build data
(train_x, train_y, test_x, test_y) = build_data(length)
print(train_x.shape, train_y.shape, test_x.shape, test_y.shape) X = tf.placeholder("float", [None, length, vector_size])
Y = tf.placeholder("float", [None, 1]) # get lstm_size and output predicted value
W = tf.Variable(tf.random_normal([10, 1], stddev=0.01))
B = tf.Variable(tf.random_normal([1], stddev=0.01)) def seq_predict_model(X, w, b, time_step_size, vector_size):
# input X shape: [batch_size, time_step_size, vector_size]
# transpose X to [time_step_size, batch_size, vector_size]
X = tf.transpose(X, [1, 0, 2]) # reshape X to [time_step_size * batch_size, vector_size]
X = tf.reshape(X, [-1, vector_size]) # split X, array[time_step_size], shape: [batch_size, vector_size]
X = tf.split(X, time_step_size, 0) # LSTM model with state_size = 10
cell = core_rnn_cell.BasicLSTMCell(num_units=10,
forget_bias=1.0,
state_is_tuple=True)
outputs, _states = core_rnn.static_rnn(cell, X, dtype=tf.float32) # Linear activation
return tf.matmul(outputs[-1], w) + b, cell.state_size pred_y, _ = seq_predict_model(X, W, B, time_step_size, vector_size)
loss = tf.square(tf.subtract(Y, pred_y))
train_op = tf.train.GradientDescentOptimizer(0.001).minimize(loss) with tf.Session() as sess:
tf.global_variables_initializer().run() # train
for i in range(50):
# train
for end in range(batch_size, len(train_x), batch_size):
begin = end - batch_size
x_value = train_x[begin: end]
y_value = train_y[begin: end]
sess.run(train_op, feed_dict={X: x_value, Y: y_value}) # randomly select validation set from test set
test_indices = np.arange(len(test_x))
np.random.shuffle(test_indices)
test_indices = test_indices[0: test_size]
x_value = test_x[test_indices]
y_value = test_y[test_indices] # eval in validation set
val_loss = np.mean(sess.run(loss,
feed_dict={X: x_value, Y: y_value}))
print('Run %s' % i, val_loss) for b in range(0, len(test_x), test_size):
x_value = test_x[b: b + test_size]
y_value = test_y[b: b + test_size]
pred = sess.run(pred_y, feed_dict={X: x_value})
for i in range(len(pred)):
print(pred[i], y_value[i], pred[i] - y_value[i])
RNN :
import random import numpy as np
import tensorflow as tf
from tensorflow.contrib.rnn.python.ops import core_rnn
from tensorflow.contrib.rnn.python.ops import core_rnn_cell def build_data(n):
xs = []
ys = []
for i in range(0, 2000):
k = random.uniform(1, 50) x = [[np.sin(k + j)] for j in range(0, n)]
y = [np.sin(k + n)] # x[i] = sin(k + i) (i = 0, 1, ..., n-1)
# y[i] = sin(k + n)
xs.append(x)
ys.append(y) train_x = np.array(xs[0: 1500])
train_y = np.array(ys[0: 1500])
test_x = np.array(xs[1500:])
test_y = np.array(ys[1500:])
return (train_x, train_y, test_x, test_y) length = 10
time_step_size = length
vector_size = 1
batch_size = 10
test_size = 10 # build data
(train_x, train_y, test_x, test_y) = build_data(length)
print(train_x.shape, train_y.shape, test_x.shape, test_y.shape) X = tf.placeholder("float", [None, length, vector_size])
Y = tf.placeholder("float", [None, 1]) # get lstm_size and output predicted value
W = tf.Variable(tf.random_normal([10, 1], stddev=0.01))
B = tf.Variable(tf.random_normal([1], stddev=0.01)) def seq_predict_model(X, w, b, time_step_size, vector_size):
# input X shape: [batch_size, time_step_size, vector_size]
# transpose X to [time_step_size, batch_size, vector_size]
X = tf.transpose(X, [1, 0, 2])
# reshape X to [time_step_size * batch_size, vector_size]
X = tf.reshape(X, [-1, vector_size])
# split X, array[time_step_size], shape: [batch_size, vector_size]
X = tf.split(X, time_step_size, 0) cell = core_rnn_cell.BasicRNNCell(num_units=10)
initial_state = tf.zeros([batch_size, cell.state_size])
outputs, _states = core_rnn.static_rnn(cell, X, initial_state=initial_state) # Linear activation
return tf.matmul(outputs[-1], w) + b, cell.state_size pred_y, _ = seq_predict_model(X, W, B, time_step_size, vector_size)
loss = tf.square(tf.subtract(Y, pred_y))
train_op = tf.train.GradientDescentOptimizer(0.001).minimize(loss) with tf.Session() as sess:
tf.global_variables_initializer().run() # train
for i in range(50):
# train
for end in range(batch_size, len(train_x), batch_size):
begin = end - batch_size
x_value = train_x[begin: end]
y_value = train_y[begin: end]
sess.run(train_op, feed_dict={X: x_value, Y: y_value}) # randomly select validation set from test set
test_indices = np.arange(len(test_x))
np.random.shuffle(test_indices)
test_indices = test_indices[0: test_size]
x_value = test_x[test_indices]
y_value = test_y[test_indices] # eval in validation set
val_loss = np.mean(sess.run(loss,
feed_dict={X: x_value, Y: y_value}))
print('Run %s' % i, val_loss) for b in range(0, len(test_x), test_size):
x_value = test_x[b: b + test_size]
y_value = test_y[b: b + test_size]
pred = sess.run(pred_y, feed_dict={X: x_value})
for i in range(len(pred)):
print(pred[i], y_value[i], pred[i] - y_value[i])
-------------------------------------------------------------------
Tensorflow 循环神经网络 基本 RNN 和 LSTM 网络 拟合、预测sin曲线的更多相关文章
- 深度学习项目——基于循环神经网络(RNN)的智能聊天机器人系统
基于循环神经网络(RNN)的智能聊天机器人系统 本设计研究智能聊天机器人技术,基于循环神经网络构建了一套智能聊天机器人系统,系统将由以下几个部分构成:制作问答聊天数据集.RNN神经网络搭建.seq2s ...
- 大话循环神经网络(RNN)
在上一篇文章中,介绍了 卷积神经网络(CNN)的算法原理,CNN在图像识别中有着强大.广泛的应用,但有一些场景用CNN却无法得到有效地解决,例如: 语音识别,要按顺序处理每一帧的声音信息,有些结果 ...
- Coursera Deep Learning笔记 序列模型(一)循环序列模型[RNN GRU LSTM]
参考1 参考2 参考3 1. 为什么选择序列模型 序列模型能够应用在许多领域,例如: 语音识别 音乐发生器 情感分类 DNA序列分析 机器翻译 视频动作识别 命名实体识别 这些序列模型都可以称作使用标 ...
- 【学习笔记】循环神经网络(RNN)
前言 多方寻找视频于博客.学习笔记,依然不能完全熟悉RNN,因此决定还是回到书本(<神经网络与深度学习>第六章),一点点把啃下来,因为这一章对于整个NLP学习十分重要,我想打好基础. 当然 ...
- TensorFlow(十一):递归神经网络(RNN与LSTM)
RNN RNN(Recurrent Neural Networks,循环神经网络)不仅会学习当前时刻的信息,也会依赖之前的序列信息.由于其特殊的网络模型结构解决了信息保存的问题.所以RNN对处理时间序 ...
- CNN(卷积神经网络)、RNN(循环神经网络)、DNN,LSTM
http://cs231n.github.io/neural-networks-1 https://arxiv.org/pdf/1603.07285.pdf https://adeshpande3.g ...
- 深度学习之循环神经网络(RNN)
循环神经网络(Recurrent Neural Network,RNN)是一类具有短期记忆能力的神经网络,适合用于处理视频.语音.文本等与时序相关的问题.在循环神经网络中,神经元不但可以接收其他神经元 ...
- TensorFlow——循环神经网络基本结构
1.导入依赖包,初始化一些常量 import collections import numpy as np import tensorflow as tf TRAIN_DATA = "./d ...
- [深度学习]理解RNN, GRU, LSTM 网络
Recurrent Neural Networks(RNN) 人类并不是每时每刻都从一片空白的大脑开始他们的思考.在你阅读这篇文章时候,你都是基于自己已经拥有的对先前所见词的理解来推断当前词的真实含义 ...
随机推荐
- Problem E: 数量的类模板
#include<iostream> #include<iomanip> #include<algorithm> using namespace std; temp ...
- pdftk - handy tool for manipulating PDF 免费的pdf合并工具
Linux pdf合并的工具 安装工具 $ sudo apt-get install pdftk 使用 $ pdftk *.pdf cat output all-in-one.pdf &&am ...
- No converter found capable of converting from type [java.lang.String] to type [java.util.Map<java.lang.String, java.lang.String>]
java.lang.IllegalStateException: Failed to load ApplicationContext at org.springframework.test.conte ...
- vscode常用快捷键总结
记住快捷键能够提高工作效率 Ctrl+Shift+P,F1 展示全局命令面板 Ctrl+P 快速打开最近打开的文件 Ctrl+Shift+N 打开新的编辑器窗口 Ctrl+Shift+W 关闭编辑器 ...
- 《你们都是魔鬼吗》第八次团队作业:第一天Alpha冲刺
<你们都是魔鬼吗>第八次团队作业:Alpha冲刺 项目 内容 这个作业属于哪个课程 任课教师博客主页链接 这个作业的要求在哪里 作业链接地址 团队名称 你们都是魔鬼吗 作业学习目标 完成最 ...
- 开源项目阅读笔记--appium+adb
git上搜了几个platform的代码,有一个项目给我感触挺深的. https://github.com/ThomasHansson/Appium-cross-platform-example/tre ...
- 为何基于tcp协议的通信比基于udp协议的通信更可靠?
tcp协议一定是先建好双向链接,发一个数据包要得到确认才算发送完成,没有收到就一直给你重发:udp协议没有链接存在,udp直接丢数据,不管你有没有收到. TCP的可靠保证,是它的三次握手双向机制,这一 ...
- C# 函数参数中的this
先看下面的代码: public static class StringExtension { public static void Foo(this string s) { Console.Write ...
- Python + Apache Kylin 让数据分析更加简单!
现如今,大数据.数据科学和机器学习不仅是技术圈的热门话题,也是当今社会的重要组成.数据就在每个人身边,同时每天正以惊人的速度快速增长,据福布斯报道:到 2025 年,每年将产生大约 175 个 Zet ...
- MongoDB 查看集合的统计信息
和 RDBMS 一样, MongoDB 同样存储集合的统计信息,通过调用命令 db.collection.stats() 可以方便的查看集合的统计信息. --1 查看集合 things 的统计信息 r ...