Keras调用LSTM进行时间序列预测（sequence-to-sequence）

keras环境搭建
#Install numpy and scipy
sudo apt-get install gfortran libopenblas-dev liblapack-dev libatlas-base-dev python-pip g++ libopenblas-dev git python-nose python-pip python-dev python-tk
sudo apt-get remove python-numpy python-scipy
sudo pip install numpy
python -c "import numpy;numpy.test()"
sudo pip install scipy
python -c "import scipy;scipy.test()"

#Install sklearn
sudo pip install sklearn

#Install pandas
sudo pip install pandas

#Install pyyaml
sudo pip install pyyaml

#Install matplotlib
sudo pip install matplotlib

#Theano Install
sudo pip install Theano

#Tenserflow without GPU
export TF_BINARY_URL=https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-0.11.0rc0-cp27-none-linux_x86_64.whl
sudo pip install --upgrade $TF_BINARY_URL

#Theano Keras
sudo pip install Keras

 cudnn配置

GPU tensorflow

Pip安装时会因为网络环境安装失败，也会缺少依赖此时可以用apt-get install 安装 例如

pip install pandas
替换为
sudo apt-get install python-pandas

tensorflow如果安装失败，可以换下面这种方法

 sudo pip install --upgrade https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-0.8.0-cp27-none-linux_x86_64.whl

Lstm介绍

http://www.cnblogs.com/arkenstone/p/5794063.html

Keras主要模块介绍及 LSTM代码分析

http://www.infocool.net/kb/Other/201610/203183.html

keras官方教程

LSTM实例

基于Theano的深度学习(Deep Learning)框架Keras学习随笔-01-FAQ

基于Theano的深度学习(Deep Learning)框架Keras学习随笔-02-Example

基于Theano的深度学习(Deep Learning)框架Keras学习随笔-05-模型

以下是广东航空大赛的一个时间序列预测代码（sequence-to-sequence）

import osfrom os import pathimport pandas as pd  from random import randomimport numpy as npimport pylab as pl
def _load_data(data, lookback = 10):  
    docX, docY = [], []
    for i in range(len(data)-lookback):#        docX.append(data.iloc[i:i+lookback].as_matrix())
        docY.append(data.iloc[i+lookback].as_matrix())
    alsX = np.array(docX)
    alsY = np.array(docY)
    #print('alsX.shape:'+str(alsX.shape))
    #print('alsY.shape:'+str(alsY.shape))
    return alsX, alsY
def train_test_split(df,lookback=10):  
    X_train, y_train = _load_data(df.iloc[0:len(df)],lookback)
    #X_test=np.array([df.iloc[len(df)-1-lookback:len(df)-1].as_matrix()])
    X_test=np.array([df.iloc[len(df)-lookback:len(df)].as_matrix()])
    print(X_test.shape)
    print(y_train.shape)
    #y_test=np.array([df.iloc[len(df)-1].as_matrix()])
    #print(y_test.shape)
    #return (X_train, y_train), (X_test, y_test)
    return (X_train, y_train), (X_test)
def find_files(path,endswith):# the return type is a list
    names=[name for name in os.listdir(path) if os.path.isfile(os.path.join(path,name)) and name.endswith(endswith)]
        if(path.endswith('/')):
                for i in range(len(names)):
            names[i]=path+names[i]
        else:
                for i in range(len(names)):
            names[i]=path+'/'+names[i]
        return names

def switch(i):
    return{0:'5-0',1:'5-1',2:'5-2',3:'5-3',4:'5-4',5:'5-5',6:'6-0',7:'6-1',8:'6-2',9:'6-3',10:'6-4',11:'6-5',12:'7-0',13:'7-1',14:'7-2',15:'7-3',16:'7-4',17:'7-5'}.get(i,[9,'float32'])
from keras.models import Sequential  from keras.layers.core import Dense, Activation  from keras.layers.recurrent import LSTM


in_neurons = 18  
out_neurons = 18
batch_size=1;#params that can be tuned  
hidden_neurons = 36
lookback=36;
nb_epoch=100;

files=find_files('./dataset_vec_new','.csv')for srcfile in files:
    data = pd.read_csv(srcfile, engine='python')
    wifiname=srcfile[18:]
        print "wifiname"
    wifiname=wifiname.replace('_10min_vec.csv','')
        print wifiname
    print('begin processing '+wifiname+'......')
    model = Sequential() 
    model.add(LSTM(output_dim=hidden_neurons, input_dim=in_neurons, return_sequences=False)) 
    model.add(Dense(input_dim=hidden_neurons,output_dim=out_neurons))
    model.add(Activation("linear"))  
    model.compile(loss="mean_squared_error", optimizer="rmsprop")  

    (X_train, y_train), (X_test) = train_test_split(data,lookback)  # retrieve data
    #print(X_test[0,9,:])
    #print(y_train[2093,:])
    model.fit(X_train, y_train, batch_size=batch_size, nb_epoch=nb_epoch, validation_split=0.05)  


    predicted = model.predict(X_test)  
    #rmse = np.sqrt(((predicted - y_test) ** 2).mean(axis=0)).mean()
    #print('RMSE:'+str(rmse))
    fobj=open('dataset_results/'+wifiname+'_results.csv','w')
    print(predicted[0,:])
    for i in range(len(predicted[0,:])):
        fobj.write(wifiname+','+str(predicted[0,i])+','+'2016-09-25-1'+switch(i)+'\n')
    fobj.close()
    print('end processing '+wifiname+'......')
#print(predicted.shape)
#print(y_test)
#print(X_test[0,9,:])
# and maybe plot it
#pd.DataFrame(predicted[:100]).to_csv("predicted.csv")  
#pd.DataFrame(y_test[:100]).to_csv("test_data.csv")  
#pl.plot(predicted[0,:],'*')
#pl.plot(y_test[0,:],'*')
#pl.show()
#pd.DataFrame(predicted[0,:]).plot(kind='dot')
#pd.DataFrame(y_test[0,:]).plot(kind='line')

数据集保存在文件中

 

转载于:https://www.cnblogs.com/squirrel-shy/p/5986094.html