如何为Keras LSTM模型正确指定输入形状

Question

I have a generator that yields my x and y which is what I'm using to train my LSMT model in Keras.

The X is of the format [0.0, 1.0, 0.0004].

I set the shape of this data earlier:

x_out = np.array(list(x_drop.values())).reshape(len(x_drop), 1)
y_out = np.array([y])

Which I get in an async generator, so then I pass into a syncing function and then use model.fit_generator():

train_gen = to_sync_generator(replay(traintime_0, traintime_1, test=0))
    model = lstm_model()

model.fit_generator(
    train_gen,
    epochs = 5,
    verbose = 1,
    steps_per_epoch=500,
    initial_epoch=0,
    validation_data=valid_gen,
    validation_steps=500)

My model code is below.

   def lstm_model():
       model = Sequential()
       model.add(LSTM(100, input_shape=(3,1)))
       model.add(Dense(1))
       model.compile(loss='mean_squared_error', optimizer='adam')
       return model

I get the following error:

ValueError: Error when checking input: expected lstm_1_input to have 3 dimensions, but got array with shape (3, 1)

Note, I can not preprocess this data.

I haven't really seen examples of using lstm with non-preprocessed data, any advise is greatly appreciated! thank you.

Answer 1

Your data must be a tuple using the format (x, y, z), where => x is your batch number (you can initially start this with 1), y is your data array and z is your data shape. By your case, I can deduce your data have one feature and has 3 steps. So, you can prepare it using the following code:

data = data.reshape((1, 3, 1))

If you cannot change the data directly, you can merge it into a new array, then fit the array into the model