使用 LSTM 自动编码器时出现 NaN 错误

Question

>我正在尝试使用 Keras 使用 LSTM 自动编码器训练模型，以重建我提供给模型的输入，结果中出现 NaN 错误，我在解码部分后获得。这是我的代码;

    # lstm autoencoder recreate sequence
    from numpy import array
    import numpy as np
    from keras.models import Sequential
    from keras.layers import LSTM
    from keras.layers import Dense
    from keras.layers import RepeatVector
    from keras.layers import TimeDistributed
    from keras.utils import plot_model
    import pandas as pd
    df = pd.read_csv('flight_data.csv',sep=',',header=None)
    data = df.to_numpy()
    print(data.shape)

    # define input sequence
    sequence1 = array([0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9])
    sequence2 = array([0.2, 0.4, 0.6, 0.4, 1.0, 1.2, 1.4, 1.6, 1.8])
    # reshape input into [samples, timesteps, features]
    n_in = 100
    data = data[73666:,:]
    sequence = data.reshape((1,100,24))
    print(sequence)
    # define model
    model = Sequential()
    model.add(LSTM(100, activation='relu', input_shape=(n_in,24)))
    model.add(RepeatVector(n_in))
    model.add(LSTM(100, activation='relu', return_sequences=True))
    model.add(TimeDistributed(Dense(24)))
    model.compile(optimizer='adam', loss='mse')
    # fit model
    model.fit(sequence, sequence, epochs=300, verbose=0)
    plot_model(model, show_shapes=True, to_file='reconstruct_lstm_autoencoder.png')
    # demonstrate recreation
    yhat = model.predict(sequence, verbose=0)
    print(yhat)

我得到的输出是;

[[[9.46687355e+14 1.00000000e+01 4.42748822e+08 ... 0.00000000e+00
   0.00000000e+00 0.00000000e+00]
  [9.46687355e+14 1.00000000e+01 4.42748822e+08 ... 0.00000000e+00
   0.00000000e+00 0.00000000e+00]
  [9.46687355e+14 1.00000000e+01 4.42748823e+08 ... 0.00000000e+00
   0.00000000e+00 0.00000000e+00]
  ...
  [9.46687359e+14 1.00000000e+01 4.42748824e+08 ... 0.00000000e+00
   0.00000000e+00 0.00000000e+00]
  [9.46687359e+14 1.00000000e+01 4.42748824e+08 ... 0.00000000e+00
   0.00000000e+00 0.00000000e+00]
  [9.46687359e+14 1.00000000e+01 4.42748825e+08 ... 0.00000000e+00
   0.00000000e+00 0.00000000e+00]]]
[[[nan nan nan ... nan nan nan]
  [nan nan nan ... nan nan nan]
  [nan nan nan ... nan nan nan]
  ...
  [nan nan nan ... nan nan nan]
  [nan nan nan ... nan nan nan]
  [nan nan nan ... nan nan nan]]]

哪个部分可能会导致问题？我该怎么办？

Answer 1

这看起来你有爆炸梯度，LSTM 有创建的趋势。裁剪渐变可以解决此问题，请尝试将裁剪范数设置为 1。

ADAM = keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False, clipnorm=1.)
model.compile(optimizer=ADAM, loss='mse')