RNN time series predictions with multiple time series dimension with Keras, Tensorflow

问题

I am trying to run a RNN/LSTM network on some time series sets. It should be mentioned that the time series are being classified. I have ~600 different time series, and each of these has 930 timesteps with features in them. I have structured my data into a numpy 3D array that is structured like:

X = [666 observations/series, 930 timesteps in each observation, 15 features]
Y = [666 observations/series, 930 timesteps in each observation, 2 features]

For training and validation data I split the data 70/30. So Train_X = [466, 930, 15] and Train_Y = [200, 930, 2].

My network is getting an error that is saying that it expected the input to be 2 dimensions and that it got an array with shape (466, 930, 2). My code is as follows:

from sklearn.preprocessing import MinMaxScaler
from sklearn.preprocessing import LabelEncoder
from sklearn.metrics import mean_squared_error

from keras.models import Sequential
from keras.layers import Dense
from keras.layers import LSTM
from keras.layers import Bidirectional

Train_X = new_ped_data[0:466]
Test_X = new_ped_data[466:]

Train_Y = new_ped_valid_data[0:466]
Test_Y = new_ped_valid_data[466:]

model = Sequential()
model.add(Bidirectional(LSTM(20, return_sequences=True),
                        input_shape=Train_X.shape[1:]))
model.add(Bidirectional(LSTM(10)))
model.add(Dense(5))
model.compile(loss='mae', 
              optimizer='rmsprop')

model.fit(Train_X, Train_Y, epochs = 30, batch_size = 32, 
      validation_data =(Test_X, Test_Y))

I am just trying to get the model running. Once I do, then I'll tweak the architecture and fit parameters. I should mention that one of the classification outputs might not be necessary. Any suggestions as to how I set up the architecture so that if a time series is fed in I will get the classification values of the network for each timestep?

Error was: ValueError: Error when checking target: expected dense_9 to
have 2 dimensions, but got array with shape (466, 930, 2)

回答1:

Your output has also sequential nature. LSTM by default has a flag return_sequences=False. This makes your sequence to be squashed to a vector after the second LSTM layer. In order to change that try:

model = Sequential()
model.add(Bidirectional(LSTM(20, return_sequences=True),
                    input_shape=Train_X.shape[1:]))
model.add(Bidirectional(LSTM(10, return_sequences=True)))
model.add(Dense(5))
model.compile(loss='mae', 
          optimizer='rmsprop')

来源：https://stackoverflow.com/questions/48253348/rnn-time-series-predictions-with-multiple-time-series-dimension-with-keras-tens