目录
1.已知知识
1.1LSTM
指长短期记忆人工神经网络。长短期记忆网络(LSTM,Long Short-Term Memory)是一种时间循环神经网络,是为了解决一般的RNN(循环神经网络)存在的长期依赖问题而专门设计出来的。
RNN:Recurrent Neural Network 循环神经网络的计算过程如下:
,
,
,
在LSTM当中将换成了,而且,对的更新也换成了。
墙裂推荐吴恩达的课程,图片来自他讲LSTM,非常清楚,LSTM的计算过程如下:
计算三个门的结果除了外还可以再加一个上次细胞状态 ,这个操作叫窥视连孔连接,peophole connection。
1.2.随机行走模型
用random walk model 得出概率的走向,下面这段代码可以玩一下。
__author__ = 'Administrator'
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
fig=plt.figure()
#time span
T=500
#drift factor飘移率
mu=0.00005
#volatility波动率
sigma=0.04
#t=0初试价
S0=np.random.random()
#length of steps
dt=1
N=round(T/dt)
t=np.linspace(0,T,N)
W=np.random.standard_normal(size=N)
print("W ",W.shape)
#W.shape=(500,)
#几何布朗运动过程
W=np.cumsum(W)*np.sqrt(dt)
X=(mu-0.5*sigma**2)*t+sigma*W
S=S0*np.exp(X)
fd=pd.DataFrame({'pro':S})
fd.to_csv('pic/random_walk.csv',sep=',',index=False)
plt.plot(t,S,lw=2)
plt.show()
2 问题描述
预测某个app在未来某个时间再次被打开的概率,其概率曲线用随机行走模型 random walk model 得出,数据大小都为0-1之间的小数,如果得出的图的取值不在【0,1】,多画几次~,因为每个动作都随机,很有可能会超过【0,1】这个范围。假设共500个分钟,先用70%的数据进行训练,打乱数据集后,再选取30%的数据进行测试,这样可以提高泛化能力。
配置:用cpu跑的,相当的慢了,使用keras
3 代码
import pandas as pd
import numpy as np
import keras
import matplotlib.pyplot as plt
#from sklearn.preprocessing import MinMaxScaler
from keras.models import Sequential
from keras.layers import LSTM, Dense, Activation
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
test_num=500
train_times=1000
#random walk model to generate the probability tendency of task
def random_walk_model():
fig=plt.figure()
#time span
T=500
#drift factor飘移率
mu=0.00005
#volatility波动率
sigma=0.04
#t=0初试价
S0=np.random.random()
#length of steps
dt=1
N=round(T/dt)
#generate 500 steps and collect it into t
t=np.linspace(0,T,N)
#W is standard normal list
W=np.random.standard_normal(size=N)
print("W ",W)
#W.shape=(500,)
#几何布朗运动过程,产生概率轨迹
W=np.cumsum(W)*np.sqrt(dt)
X=(mu-0.5*sigma**2)*t+sigma*W
S=S0*np.exp(X)
plt.plot(t,S,lw=2)
plt.show()
#save the probability tendency of picture
fd=pd.DataFrame({'pro':S})
fd.to_csv('pic/random_walk_model.csv',sep=',',index=False)
plt.savefig('pic/random_data.png')
return S
def random_test(sequence_length=5,split=0.7):
#get the stored data by using pandas
test_data = pd.read_csv('pic/random_walk_model.csv', sep=',',usecols=[0])
#print("test_data:",test_data)
#generate new test data for 2d
test_data = np.array(test_data).astype('float64')
#print('test_data:',test_data.shape)
#test_data: (500, 1)
#70% are used to be trained, the rest is used to be tested
split_boundary = int(test_data.shape[0] * split)
#print('split_boundary:',split_boundary)
#split_boundary:350
pro_test=np.linspace(split_boundary,test_data.shape[0],test_data.shape[0]-split_boundary)
pro_x=np.linspace(1,split_boundary,split_boundary)
plt.plot(pro_x,test_data[:split_boundary])
plt.plot(pro_test,test_data[split_boundary:],'red')
plt.legend(['train_data','test_data'])
plt.xlabel('times')
plt.ylabel('probability')
plt.show()
#print("test_data: ",test_data,test_data.shape),test_data.shape=(600,1),array to list format
#generate 3d format of data and collect it
data = []
for i in range(len(test_data) - sequence_length - 1):
data.append(test_data[i: i + sequence_length + 1])
#print(len(data[0][0]),len(data[1]),len(data))
#1 6 494
reshaped_data = np.array(data).astype('float64')
#print("reshaped_data:",reshaped_data.shape)
#reshaped_data: (494, 6, 1)
#random the order of test_data to improve the robustness
np.random.shuffle(reshaped_data)
#from n to n*5 are the training data collected in x, the n*6th is the true value collected in y
x = reshaped_data[:, :-1]
y = reshaped_data[:, -1]
#print("x ",x.shape,"\ny ",y.shape)
#x (494, 5, 1) y (494, 1)
#train data
train_x = x[: split_boundary]
train_y = y[: split_boundary]
#test data
test_x = x[split_boundary:]
test_y=y[split_boundary:]
#print("train_y:",train_x.shape,"train_y:",train_y.shape,"test_x ",test_x.shape,"test_y",test_y.shape)
#train_y: (350, 5, 1) train_y: (350, 1) test_x (144, 5, 1) test_y (144, 1)
return train_x, train_y, test_x, test_y
def build_model():
# input_dim是输入的train_x的最后一个维度,相当于输入的神经只有1个——特征只有1个,train_x的维度为(n_samples, time_steps, input_dim)
#如果return_sequences=True:返回形如(samples,timesteps,output_dim)的3D张量否则,返回形如(samples,output_dim)的2D张量
#unit并不是输出的维度,而是门结构(forget门、update门、output门)使用的隐藏单元个数
model = Sequential()
#use rmsprop for optimizer
rmsprop=keras.optimizers.RMSprop(lr=0.001, rho=0.9,epsilon=1e-08,decay=0.0)
#build one LSTM layer
model.add(LSTM(input_dim=1, units=1, return_sequences=False,use_bias=True,activation='tanh'))
#model.add(LSTM(100, return_sequences=False,use_bias=True,activation='tanh'))
#comiple this model
model.compile(loss='mse', optimizer=rmsprop)#rmsprop
return model
def train_model(train_x, train_y, test_x, test_y):
#call function to build model
model = build_model()
try:
#store this model to use its loss parameter
history=model.fit(train_x, train_y, batch_size=20, epochs=train_times,verbose=2)
#store the loss
lossof_history=history.history['loss']
predict = model.predict(test_x)
predict = np.reshape(predict, (predict.size, ))
#evaluate this model by returning a loss
loss=model.evaluate(test_x,test_y)
print("loss is ",loss)
#if there is a KeyboardInterrupt error, do the following
except KeyboardInterrupt:
print("error of predict ",predict)
print("error of test_y: ",test_y)
try:
#x1 is the xlabel to print the test value, there are 500 data,30% is for testing
x1=np.linspace(1,test_y.shape[0],test_y.shape[0])
#x1 is the xlabel to print the loss value, there are 500 data,70% is for training
x2=np.linspace(1,train_times,train_times)
fig = plt.figure(1)
#print the predicted value and true value
plt.title("test with rmsprop lr=0.01_")
plt.plot(x1,predict,'ro-')
plt.plot(x1,test_y,'go-')
plt.legend(['predict', 'true'])
plt.xlabel('times')
plt.ylabel('propability')
plt.savefig('pic/train_with_rmsprop_lr=0.01.png')
#print the loss
fig2=plt.figure(2)
plt.title("loss lr=0.01")
plt.plot(x2,lossof_history)
plt.savefig('pic/train_with_rmsprop_lr=0.01_LOSS_.png')
plt.show()
#if the len(x1) is not equal to predict.shape[0] / test_y.shape[0] / len(x2) is not equal to lossof_history.shape[0],there will be an Exception
except Exception as e:
print("error: ",e)
if __name__ == '__main__':
#random_walk_model() function is only used by once, because data are stored as pic/random_data.csv
#random_walk_model()
#prepare the right data format for LSTM
train_x, train_y, test_x, test_y=random_test()
#standard the format for LSTM input
test_x = np.reshape(test_x, (test_x.shape[0], test_x.shape[1], 1))
#print("main: train_x.shape ",train_x.shape)
#main: train_x.shape (350, 5, 1)
train_model(train_x, train_y, test_x, test_y)
细节
3.1.数据准备
def random_test(sequence_length=6,split=0.7):
在这个函数当中,将获取的数据处理为需要的三维的张量格式,在训练的时候
history=model.fit(train_x, train_y, batch_size=20, epochs=train_times,verbose=2)
train_x.shape= (415, 6, 1)#600条数据,70%的训练数据,30%测试数据
最后使用的训练数据需要是这样的三维张量。
Question:600条的数据,70%用于训练,那么数据应该是420条,但是为什么是415?
Answer:在预处理数据之后,建立一层的LSTM:
model.add(LSTM(input_dim=1, units=50, return_sequences=False))
后台显示input_dim warning,找了下原因,要使用input_shape作为参数,input_shape为3维,参数为(Batch_size, Time_step, Input_Sizes)。
batch_size设置:batch_size为
在model.add(LSTM(input_shape(,,,), units=, return_sequences=))语句中定义 | 在该语句中不定义batch_size,input_shape=(5,),这只定义了time_step,或者input_shape=(None,5,5),第一个参数定义为None |
无法使用model.train_on_batch(),且在test时也需要有batch_size的数据 | 可以调用train_on_batch() |
time_step为时间序列的长度/语句的最大长度
input_sizes:每个时间点输入x的维度/语句的embedding的向量维度,本例题做概率预测,再次回忆LSTM的图,x(t)输入是一个概率值,那么input_sizes=1,即特征值只有一个
那这与数据415条的关系在哪呢?
注意默认的参数 sequence_length=6
这是数据的一部分:
'''
0.7586717205211277
0.6628550358816061
0.9184003785782959
0.09365662435769384
0.9791582266747239
0.8700739252039772
0.7924134549615585
0.3983410609045436
0.38988445126231197
0.8167186985712294
0.879351951255656
0.9468282424096985
0.7060727836006101
0.7650727081508003
0.3633755461129521
0.3489589275449808
'''
for i in range(len(test_data) - sequence_length - 1):
data.append(test_data[i: i + sequence_length + 1])
在上面的语句作用就是
当i=0时,data.append(test[0:6]),一共循环594次,那么data的数据为
[[[0.75867172]
[0.66285504]
[0.91840038]
[0.09365662]
[0.97915823]
[0.87007393]]
[[0.66285504]
[0.91840038]
[0.09365662]
[0.97915823]
[0.87007393]
[0.79241345]]
[[0.91840038]
[0.09365662]
[0.97915823]
[0.87007393]
[0.79241345]
[0.39834106]]
...
]
即 其步长为1,生成的data.shape=(594, 6, 1)= (seq_len, batch, input_size)
其中第一个参数表示数据个数,
第二个参数表示,观察dada的数据,步长为1(自己设定的),个数为6,进行划分,batch就是那个6
第三个参数:回忆上面讲的input_shape的input_size——>每个时间点输入x的维度/语句的embedding的向量维度,本例题做概率预测,再次回忆LSTM的图,x(t)输入是一个概率值,那么input_sizes=1,即特征值只有一个
3.2.结果
用于测试的30%的数据预测结果与真实结果如下