学习了RNN和LSTM的理论知识，下面再来使用Keras实现一下这些模型。

理论知识：

Keras实现神经网络：

Keras的安装过程：

1. 环境准备

import matplotlib.pyplot as plt
from math import sqrt
from matplotlib import pyplot
import pandas as pd
import numpy as np
from numpy import concatenate
from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation
from keras import layers
from keras.optimizers import Adam

2. 加载数据及数据预处理

数据集为26个特征变量及一个预测值。数据集总共14122条数据，其中训练集取前15年数据，共10066条，后5年做预测，共4055条。

将数据集进行划分，然后将训练集和测试集划分为输入和输出变量，最终将输入（X）改造为RNN，LSTM的输入格式，即[samples,timesteps,features]。

filename = 'data.csv'
data = pd.read_csv(filename)

values = data.values
# # 原始数据标准化，为了加速收敛
scaler = MinMaxScaler(feature_range=(0, 1))
scaled = scaler.fit_transform(values)
Y = scaled[:, -1]
X = scaled[:, 0:-1]

train_x = X[0: 10066]
train_y = Y[0: 10066]
test_x = X[10066: 14122]
test_y = Y[10066: 14122]

train_x, train_y = np.array(train_x), np.array(train_y)
test_x, test_y = np.array(test_x), np.array(test_y)

"""
输入要求：[送入样本数， 循环核时间展开步数， 每个时间步输入特征个数]
"""
train_x = np.reshape(train_x, (train_x.shape[0], 1, 26))
test_x = np.reshape(test_x,  (test_x.shape[0], 1, 26))

我的数据集，由于每年的数据不一样，无法按年划分，没有时间序列性。因此我的循环核时间展开步数为1。

一般的时间序列，如使用60天的股票预测第61天的股票，他的时间序列就是60。具体可参考：深度学习100例-循环神经网络（RNN）实现股票预测

"""
使用前60天的开盘价作为输入特征x_train
    第61天的开盘价作为输入标签y_train
    
for循环共构建2426-300-60=2066组训练数据。
       共构建300-60=260组测试数据
"""
for i in range(60, len(training_set)):
    x_train.append(training_set[i - 60:i, 0])
    y_train.append(training_set[i, 0])
    
for i in range(60, len(test_set)):
    x_test.append(test_set[i - 60:i, 0])
    y_test.append(test_set[i, 0])

x_train, y_train = np.array(x_train), np.array(y_train) # x_train形状为：(2066, 60, 1)
x_test,  y_test  = np.array(x_test),  np.array(y_test)

"""
输入要求：[送入样本数， 循环核时间展开步数， 每个时间步输入特征个数]
"""
x_train = np.reshape(x_train, (x_train.shape[0], 60, 1))
x_test  = np.reshape(x_test,  (x_test.shape[0], 60, 1))

其中，在所有的RNN中，包括simpleRNN, LSTM, GRU等等，输入输出数据格式如下：
在这里插入图片描述

2. 构建、编译模型

RNN模型是调用layers.SimpleRNN，LSTM模型是调用layers.LSTM

RNN模型：

# 创建模型
model = Sequential()
# RNN神经网络
# 隐藏层100
model.add(layers.SimpleRNN(units=100, return_sequences=True))
model.add(Activation('relu'))
# Dropout层用于防止过拟合
model.add(Dropout(0.1))
# 隐藏层100
model.add(layers.SimpleRNN(units=100))
model.add(Activation('relu'))
model.add(Dropout(0.1))
model.add(Dense(1))

LSTM模型：

# 创建模型
model = Sequential()
# LSTM神经网络
# 隐藏层100
model.add(layers.LSTM(units=100, activation='relu', return_sequences=True))
# 隐藏层100
model.add(layers.LSTM(units=100, activation='relu'))
model.add(Dense(1))

# 编译模型
# 使用高效的ADAM优化算法以及优化的最小均方误差损失函数
model.compile(loss='mean_squared_error', optimizer=Adam())

3. 训练模型

# 训练模型
history = model.fit(train_x, train_y, epochs=30, batch_size=64,
                    validation_data=(test_x, test_y), verbose=2)
model.summary()  # 查看你的神经网络的架构和参数量等信息

4. 模型预测及可视化

# loss曲线
pyplot.plot(history.history['loss'], label='train')
pyplot.plot(history.history['val_loss'], label='test')
pyplot.title('Model loss')
pyplot.ylabel('Loss')
pyplot.xlabel('Epoch')
plt.title('Training and Validation Loss')
pyplot.legend()
pyplot.show()

# 预测
yhat = model.predict(test_x)

test_x = np.reshape(test_x,  (test_x.shape[0], 26))


# 预测y 逆标准化
inv_yhat0 = concatenate((test_x, yhat), axis=1)
inv_yhat1 = scaler.inverse_transform(inv_yhat0)
inv_yhat = inv_yhat1[:, -1]



# 原始y逆标准化
test_y = test_y.reshape(len(test_y), 1)
inv_y0 = concatenate((test_x, test_y), axis=1)
inv_y1 = scaler.inverse_transform(inv_y0)
inv_y = inv_y1[:, -1]

print(inv_y)
# 计算 R2
r_2 = r2_score(inv_y, inv_yhat)
print('Test r_2: %.3f' % r_2)
# 计算MAE
mae = mean_absolute_error(inv_y, inv_yhat)
print('Test MAE: %.3f' % mae)
# 计算RMSE
rmse = sqrt(mean_squared_error(inv_y, inv_yhat))
print('Test RMSE: %.3f' % rmse)
plt.plot(inv_y)
plt.plot(inv_yhat)
plt.title('Intensity Prediction')
plt.show()

5. 预测结果

Train on 10066 samples, validate on 4055 samples
Epoch 1/30
 - 1s - loss: 0.0438 - val_loss: 0.0046
Epoch 2/30
 - 0s - loss: 0.0070 - val_loss: 0.0047
Epoch 3/30
 - 0s - loss: 0.0057 - val_loss: 0.0036
Epoch 4/30
 - 0s - loss: 0.0050 - val_loss: 0.0042
Epoch 5/30
 - 0s - loss: 0.0046 - val_loss: 0.0037
Epoch 6/30
 - 0s - loss: 0.0044 - val_loss: 0.0036
Epoch 7/30
 - 0s - loss: 0.0042 - val_loss: 0.0036
Epoch 8/30
 - 0s - loss: 0.0041 - val_loss: 0.0037
Epoch 9/30
 - 0s - loss: 0.0040 - val_loss: 0.0034
Epoch 10/30
 - 0s - loss: 0.0040 - val_loss: 0.0032
Epoch 11/30
 - 0s - loss: 0.0039 - val_loss: 0.0030
Epoch 12/30
 - 0s - loss: 0.0038 - val_loss: 0.0030
Epoch 13/30
 - 0s - loss: 0.0037 - val_loss: 0.0033
Epoch 14/30
 - 0s - loss: 0.0036 - val_loss: 0.0031
Epoch 15/30
 - 0s - loss: 0.0036 - val_loss: 0.0030
Epoch 16/30
 - 0s - loss: 0.0035 - val_loss: 0.0034
Epoch 17/30
 - 0s - loss: 0.0034 - val_loss: 0.0029
Epoch 18/30
 - 0s - loss: 0.0033 - val_loss: 0.0037
Epoch 19/30
 - 0s - loss: 0.0033 - val_loss: 0.0029
Epoch 20/30
 - 0s - loss: 0.0033 - val_loss: 0.0029
Epoch 21/30
 - 0s - loss: 0.0031 - val_loss: 0.0029
Epoch 22/30
 - 0s - loss: 0.0032 - val_loss: 0.0028
Epoch 23/30
 - 0s - loss: 0.0031 - val_loss: 0.0029
Epoch 24/30
 - 0s - loss: 0.0030 - val_loss: 0.0033
Epoch 25/30
 - 0s - loss: 0.0030 - val_loss: 0.0029
Epoch 26/30
 - 0s - loss: 0.0030 - val_loss: 0.0029
Epoch 27/30
 - 0s - loss: 0.0029 - val_loss: 0.0029
Epoch 28/30
 - 0s - loss: 0.0028 - val_loss: 0.0028
Epoch 29/30
 - 0s - loss: 0.0029 - val_loss: 0.0028
Epoch 30/30
 - 0s - loss: 0.0028 - val_loss: 0.0028
Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
simple_rnn_1 (SimpleRNN)     (None, 1, 100)            12700     
_________________________________________________________________
activation_1 (Activation)    (None, 1, 100)            0         
_________________________________________________________________
dropout_1 (Dropout)          (None, 1, 100)            0         
_________________________________________________________________
simple_rnn_2 (SimpleRNN)     (None, 100)               20100     
_________________________________________________________________
activation_2 (Activation)    (None, 100)               0         
_________________________________________________________________
dropout_2 (Dropout)          (None, 100)               0         
_________________________________________________________________
dense_1 (Dense)              (None, 1)                 101       
=================================================================
Total params: 32,901
Trainable params: 32,901
Non-trainable params: 0
_________________________________________________________________
[ 3.  3.  0. ... -5. -2.  0.]
Test r_2: 0.566
Test MAE: 2.452
Test RMSE: 3.371

Process finished with exit code 0

在这里插入图片描述
在这里插入图片描述

6. Keras实现RNN全部代码：

我使用RNN来做普通的回归预测，并没有时间序列，可以简单的参考一下。如果想参考时间序列的回归预测，可以参考上面的股票预测。

import matplotlib.pyplot as plt
from math import sqrt
from matplotlib import pyplot
import pandas as pd
import numpy as np
from numpy import concatenate
from sklearn.preprocessing import MinMaxScaler
from sklearn.metrics import mean_absolute_error, mean_squared_error, r2_score
from keras.models import Sequential
from keras.layers.core import Dense, Dropout, Activation
from keras import layers
from keras.optimizers import Adam

'''Keras实现神经网络回归模型'''
# 读取数据

filename = 'data.csv'
data = pd.read_csv(filename)

values = data.values
# # 原始数据标准化，为了加速收敛
scaler = MinMaxScaler(feature_range=(0, 1))
scaled = scaler.fit_transform(values)
Y = scaled[:, -1]
X = scaled[:, 0:-1]

train_x = X[0: 10066]
train_y = Y[0: 10066]
test_x = X[10066: 14122]
test_y = Y[10066: 14122]

input = X.shape[1]
time_steps = input

train_x, train_y = np.array(train_x), np.array(train_y)
test_x, test_y = np.array(test_x), np.array(test_y)

print(train_x.shape)
"""
输入要求：[送入样本数， 循环核时间展开步数， 每个时间步输入特征个数]
"""
train_x = np.reshape(train_x, (train_x.shape[0], 1, 26))
test_x = np.reshape(test_x,  (test_x.shape[0], 1, 26))

# 创建模型
model = Sequential()
# 循环神经网络
# 隐藏层100
model.add(layers.SimpleRNN(units=100, return_sequences=True))
model.add(Activation('relu'))
# Dropout层用于防止过拟合
model.add(Dropout(0.1))
# 隐藏层100
model.add(layers.SimpleRNN(units=100))
model.add(Activation('relu'))
model.add(Dropout(0.1))
model.add(Dense(1))

# 编译模型
# 使用高效的ADAM优化算法以及优化的最小均方误差损失函数
model.compile(loss='mean_squared_error', optimizer=Adam())

# 训练模型
history = model.fit(train_x, train_y, epochs=30, batch_size=64,
                    validation_data=(test_x, test_y), verbose=2)
model.summary()  # 查看你的神经网络的架构和参数量等信息

# loss曲线
pyplot.plot(history.history['loss'], label='train')
pyplot.plot(history.history['val_loss'], label='test')
pyplot.title('Model loss')
pyplot.ylabel('Loss')
pyplot.xlabel('Epoch')
plt.title('Training and Validation Loss')
pyplot.legend()
pyplot.show()

# 预测
yhat = model.predict(test_x)

test_x = np.reshape(test_x,  (test_x.shape[0], 26))


# 预测y 逆标准化
inv_yhat0 = concatenate((test_x, yhat), axis=1)
inv_yhat1 = scaler.inverse_transform(inv_yhat0)
inv_yhat = inv_yhat1[:, -1]



# 原始y逆标准化
test_y = test_y.reshape(len(test_y), 1)
inv_y0 = concatenate((test_x, test_y), axis=1)
inv_y1 = scaler.inverse_transform(inv_y0)
inv_y = inv_y1[:, -1]

print(inv_y)
# 计算 R2
r_2 = r2_score(inv_y, inv_yhat)
print('Test r_2: %.3f' % r_2)
# 计算MAE
mae = mean_absolute_error(inv_y, inv_yhat)
print('Test MAE: %.3f' % mae)
# 计算RMSE
rmse = sqrt(mean_squared_error(inv_y, inv_yhat))
print('Test RMSE: %.3f' % rmse)
plt.plot(inv_y)
plt.plot(inv_yhat)
plt.title('Intensity Prediction')
plt.show()

参考：

LSTM梳理，理解，和keras实现 （一）

深度学习100例-循环神经网络（RNN）实现股票预测

深度学习100例-循环神经网络（LSTM）实现股票预测

# keras # lstm # rnn # python
Logo
华为云开发者联盟

为开发者提供学习成长、分享交流、生态实践、资源工具等服务，帮助开发者快速成长。

更多推荐

cover

详解数仓的向量化执行引擎

avatar 华为云开发者联盟
cover

一次故障演练，十分钟自动搞定?

avatar 华为云开发者联盟
cover

实践探讨Python如何进行异常处理与日志记录

avatar 华为云开发者联盟