用于Keras LSTM模型的batch_input_shape

Question

I am trying to build a neural network with an LSTM as first hidden layer with the Keras library (tensorflow backend). I am having problem understanding how to reshape my data and feed it to a stateful LSTM with the batch_input_size parameter.

My input is a 100 second time series sampled at 10 Khz. So basically, I have 100*10000 different values of time. I have 3 different observables being sampled, so the number of features is 3. Let us call X the input matrix, with shape:

np.shape(X) = (1000000,1,3)

My target has one value per each point in time --> 100*10000 values:

np.shape(Y) = (1000000,1,1)

I want my model to take one second at a time, and therefore predict 10000 target values. I suppose this should be the batch_size . target values Say I want to put 32 nodes in the LSTM layer, this is what I wrote:

model_LSTM = Sequential()
model_LSTM.add(LSTM(32, stateful=True, batch_input_shape=(10000,1,3)))
model_LSTM.compile(optimizer=keras.optimizers.Adam(lr=0.00039, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False), loss='mean_squared_error')

To fit the model I feed it one batch at a time:

batch_size=10000
for i in range(int(X.shape[0] / batch_size)):
    X = X_l[(i * batch_size):((i + 1) * batch_size)][:][:]
    Y = Y_l[(i * batch_size):((i + 1) * batch_size)]
    model_hist = model_LSTM.fit(X, Y, epochs=1, 
                                batch_size=batch_size, 
                                verbose=1, shuffle=False)

Am I doing this correctly? The script runs without errors, but when using model_LSTM.predict() on a new set it just outputs the same value for each time step.

Answer 1

Can you provide more information such as accuracy and cost? Since, it is predicting only one class all the time, your model is not learning. Perhaps, your dataset is skewed that results in high accuracy with no learning. I would recommend tinkering hyperparameters and try different models such as SimpleRNN and GRU .

Also, since you asked I would recommend using the fit_generator method as it would decrease the overhead of calling fit again and again.