Question

我正在尝试学习rnn预测。在此示例中，我首先反馈输出。下面的示例设置输入和权重，计算输出，将数据拆分为训练/测试，将模型与训练拟合，然后预测测试。由于未保存/设置状态，因此测试预测不正确。如何保存rnn状态并进行预测。另外，如何添加其他隐藏状态？

import numpy as np
import pandas as pd
import recurrentshop

from keras.layers import Dense, Concatenate, Input
from keras.models import Model

def main():
input_data = pd.DataFrame(data={
    'A': [1, 1, -19, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 4],
    'B': [1, 1, -18, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 4]
})

depth = input_data.shape[1]

x_t = Input(shape=(depth,))  # The input to the RNN at time t
y_tm1 = Input(shape=(depth,))  # Previous output
h_tm1 = Input(shape=(depth,))

# Compute new hidden state
x_t_and_y_tm1 = Concatenate()([x_t, y_tm1])
h_t = Dense(depth, kernel_initializer='ones')(x_t_and_y_tm1)

# Build the RNN
rnn = recurrentshop.RecurrentModel(input=x_t, initial_states=[h_tm1], output=h_t, final_states=[h_t],
                                   readout_input=y_tm1, return_sequences=True)

# Build a Keras Model using our RNN layer
n_time_steps = input_data.shape[0]
x = Input(shape=(None, depth))
y = rnn(x)
model = Model(x, y)

# Run the RNN over a sequence
expanded_input_data = np.expand_dims(input_data, axis=0)
out = model.predict(expanded_input_data)

print(f"expanded_input_data: {expanded_input_data}")
print(f"out: {out}")

# now train over in/out
n_training = int(0.8*n_time_steps)

training_in, testing_in = expanded_input_data[:, :n_training, :], expanded_input_data[:, n_training:, :]
training_out, testing_out = out[:, :n_training, :], out[:, n_training:, :]

# reinitialize weights

for layer in model.layers:
    if hasattr(layer, 'kernel_initializer'):
        layer.kernel.initializer.run()
    if hasattr(layer, 'bias_initializer'):
        layer.bias.initializer.run()

model.compile(loss='mse', optimizer='adam')
model.fit(training_in, training_out, validation_data=(training_in, training_out))
predict_out = model.predict(testing_in)

print(f"testing_out: {testing_out}")
print(f"predict_out: {predict_out}")

if __name__ == '__main__':
main()

预期输出为

testing_out: [[[-9.6468736e+07 -9.6468736e+07]
  [-1.9293747e+08 -1.9293747e+08]
  [-3.8587494e+08 -3.8587494e+08]
  [-7.7174989e+08 -7.7174989e+08]
  [-1.5434998e+09 -1.5434998e+09]
  [-3.0869996e+09 -3.0869996e+09]]]

实际输出是

predict_out: [[[  2.   2.]
  [  6.   6.]
  [ 14.  14.]
  [ 30.  30.]
  [ 62.  62.]
  [132. 132.]]]

如何为rnn模型设置状态

0 个答案: