Keras LSTM input dimension setting

Question

I was trying to train a LSTM model using keras but I think I got something wrong here.

I got an error of

ValueError: Error when checking input: expected lstm_17_input to have 3 dimensions, but got array with shape (10000, 0, 20)

while my code looks like

model = Sequential()
model.add(LSTM(256, activation="relu", dropout=0.25, recurrent_dropout=0.25, input_shape=(None, 20, 64)))
model.add(Dense(1, activation="sigmoid"))
model.compile(loss='binary_crossentropy',
          optimizer='adam',
          metrics=['accuracy'])
model.fit(X_train, y_train,
      batch_size=batch_size,
      epochs=10)

where X_train has a shape of (10000, 20) and the first few data points are like

array([[ 0,  0,  0, ..., 40, 40,  9],
   [ 0,  0,  0, ..., 33, 20, 51],
   [ 0,  0,  0, ..., 54, 54, 50],
...

and y_train has a shape of (10000, ), which is a binary (0/1) label array.

Could someone point out where I was wrong here?

Answer 1

For the sake of completeness, here's what's happened.

First up, LSTM, like all layers in Keras, accepts two arguments: input_shape and batch_input_shape. The difference is in convention that input_shape does not contain the batch size, while batch_input_shape is the full input shape including the batch size.

Hence, the specification input_shape=(None, 20, 64) tells keras to expect a 4-dimensional input, which is not what you want. The correct would have been just (20,).

But that's not all. LSTM layer is a recurrent layer, hence it expects a 3-dimensional input (batch_size, timesteps, input_dim). That's why the correct specification is input_shape=(20, 1) or batch_input_shape=(10000, 20, 1). Plus, your training array should also be reshaped to denote that it has 20 time steps and 1 input feature per each step.

Hence, the solution:

X_train = np.expand_dims(X_train, 2)  # makes it (10000,20,1)
...
model = Sequential()
model.add(LSTM(..., input_shape=(20, 1)))