tensorflow实现简单RNN

使用简单RNN预测谷歌的股票

实验数据

import numpy as np
import tensorflow.keras as keras
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.preprocessing import MinMaxScaler


def RNN(x_train, y_train):
    regressor = keras.Sequential()
    # add the first RNN and the Dropout to prevent overfitting
    regressor.add(keras.layers.SimpleRNN(units=50, return_sequences=True,
                                         input_shape=(x_train.shape[1], 1)))
    # if the return_sequences is True, it means that it has another rnn upward,and it return
    # (batch_size,time_step,units) shape matrix
    # otherwise it return (batch_size,units) matrix
    regressor.add(keras.layers.Dropout(0.2))

    # add the second one
    regressor.add(keras.layers.SimpleRNN(units=50, activation='tanh', return_sequences=True))
    regressor.add(keras.layers.Dropout(0.2))

    regressor.add(keras.layers.SimpleRNN(units=50))
    regressor.add(keras.layers.Dropout(0.2))
    # add the output layer
    regressor.add(keras.layers.Dense(units=1))
    # compile
    regressor.compile(optimizer='adam', loss='mean_squared_error', metrics=['accuracy'])
    regressor.fit(x=x_train, y=y_train, epochs=30, batch_size=32)
    # print(regressor.summary())
    return regressor


def visualization(real, pred):
    plt.figure(figsize=(8, 4), dpi=80, facecolor='w', edgecolor='k')
    plt.plot(real, color="orange", label="Real value")
    plt.plot(pred, color="c", label="RNN predicted result")
    plt.legend()
    plt.xlabel("Days")
    plt.ylabel("Values")
    plt.grid(True)
    plt.show()


if __name__ == "__main__":
    data = pd.read_csv(r"dataset\geogle_stock_price\archive\Google_Stock_Price_Train.csv")
    data = data.loc[:, ["Open"]].values

    train = data[:len(data) - 50]
    test = data[len(train):]
    train.reshape(train.shape[0], 1)
    scaler = MinMaxScaler(feature_range=(0, 1))
    train_scaled = scaler.fit_transform(train)

    # plt.plot(train_scaled)
    # plt.show()
    X_train = []
    Y_train = []

    time_step = 50
    for i in range(time_step, train_scaled.shape[0]):
        X_train.append(train_scaled[i - time_step:i, 0])
        Y_train.append(train_scaled[i, 0])

    X_train, Y_train = np.array(X_train), np.array(Y_train)
    X_train = X_train.reshape(X_train.shape[0], X_train.shape[1], 1)

    inputs = data[len(data) - len(test) - time_step:]
    inputs = scaler.transform(inputs)
    X_test = []
    for i in range(time_step, inputs.shape[0]):
        X_test.append(inputs[i - time_step:i, 0])
    X_test = np.array(X_test)
    X_test = X_test.reshape(X_test.shape[0], X_test.shape[1], 1)

    model = RNN(X_train, Y_train)
    pred = model.predict(X_test)
    pred = scaler.inverse_transform(pred)

    visualization(real=test, pred=pred)
    

• 简单的RNN会遇到例如梯度消失无法嵌套太多层、单项传播无法考虑后面对前面的影响等各种问题

LSTM模型预测

# -*-coding = utf-8
import numpy as np
import tensorflow.keras as keras
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.preprocessing import MinMaxScaler


def RNN(x_train, y_train):
    regressor = keras.Sequential()
    # add the first RNN and the Dropout to prevent overfitting
    regressor.add(keras.layers.SimpleRNN(units=50, return_sequences=True,
                                         input_shape=(x_train.shape[1], 1)))
    # if the return_sequences is True, it means that it has another rnn upward,and it return
    # (batch_size,time_step,units) shape matrix
    # otherwise it return (batch_size,units) matrix
    regressor.add(keras.layers.Dropout(0.2))

    # add the second one
    regressor.add(keras.layers.SimpleRNN(units=50, activation='tanh', return_sequences=True))
    regressor.add(keras.layers.Dropout(0.2))

    regressor.add(keras.layers.SimpleRNN(units=50))
    regressor.add(keras.layers.Dropout(0.2))
    # add the output layer
    regressor.add(keras.layers.Dense(units=1))
    # compile
    regressor.compile(optimizer='adam', loss='mean_squared_error')
    # a important point: we do not give a metric such as accuracy in the CNN or ohter neural network
    #
    regressor.fit(x=x_train, y=y_train, epochs=30, batch_size=32)
    # print(regressor.summary())
    return regressor


def LSTM(x_train, y_train):
    model = keras.Sequential()
    model.add(keras.layers.LSTM(units=10, activation='tanh', input_shape=(None, 1)))
    # ten units LSTM
    model.add(keras.layers.Dense(1))
    model.compile(loss='mean_squared_error', optimizer='adam', metrics=['accuracy'])
    model.fit(x_train, y_train, batch_size=1, epochs=15)
    return model


def visualization(real, pred):
    plt.figure(figsize=(8, 4), dpi=80, facecolor='w', edgecolor='k')
    plt.plot(real, color="orange", label="Real value")
    # plt.plot(pred, color="c", label="RNN predicted result")
    plt.plot(pred, color='r', label='LSTM predicted result')
    plt.legend()
    plt.xlabel("Days")
    plt.ylabel("Values")
    plt.grid(True)
    plt.show()


if __name__ == "__main__":
    data = pd.read_csv(r"dataset\geogle_stock_price\archive\Google_Stock_Price_Train.csv")
    data = data.loc[:, ["Open"]].values

    train = data[:len(data) - 50]
    test = data[len(train):]
    train.reshape(train.shape[0], 1)
    scaler = MinMaxScaler(feature_range=(0, 1))
    train_scaled = scaler.fit_transform(train)

    # plt.plot(train_scaled)
    # plt.show()
    X_train = []
    Y_train = []

    time_step = 50
    for i in range(time_step, train_scaled.shape[0]):
        X_train.append(train_scaled[i - time_step:i, 0])
        Y_train.append(train_scaled[i, 0])

    X_train, Y_train = np.array(X_train), np.array(Y_train)
    X_train = X_train.reshape(X_train.shape[0], X_train.shape[1], 1)

    inputs = data[len(data) - len(test) - time_step:]
    inputs = scaler.transform(inputs)
    X_test = []
    for i in range(time_step, inputs.shape[0]):
        X_test.append(inputs[i - time_step:i, 0])
    X_test = np.array(X_test)
    X_test = X_test.reshape(X_test.shape[0], X_test.shape[1], 1)

    # model = RNN(X_train, Y_train)
    model = LSTM(X_train, Y_train)
    pred = model.predict(X_test)
    pred = scaler.inverse_transform(pred)

    visualization(real=test, pred=pred)

可以看到LSTM模型的预测效果比简单CNN要好很多，基本上复现了真实的股价，只是在时间上面有一些延迟。

modifie RNN