How to connect two LSTM models in Keras

Question

I want to create a model with two LSTM layers with Keras. The following code however, generates an error:

from keras.models import Sequential
from keras.layers import LSTM, Dropout, Activation
from keras.callbacks import ModelCheckpoint
from keras.utils import to_categorical

model = Sequential()
model.add(LSTM(5, activation="softmax"))
model.add(LSTM(5, activation="softmax"))

model.compile(loss='categorical_crossentropy', 
              optimizer='adam', 
              metrics=['categorical_accuracy'])

# These values are to be predicted.
directions = [-2, -1, 0, 1, 2]

# Sample data. We have three time steps, one 
# feature per timestep, and one resulting value.
data = [[[[1], [2], [3]], -1], 
        [[[3], [2], [1]], 2], 
        [[[4], [5], [7]], 1],
        [[[1], [-1], [10]], -2]]

X = []
y_ = []

# Now we take 10000 samples from the data above.
for i in np.random.choice(len(data), 10000):
    X.append(data[i][0])
    y_.append(data[i][1])

X = np.array(X)
y_ = np.array(y_)
y = to_categorical(y_ + 2, num_classes=5)

model.fit(X, y, 
          epochs=3,
          validation_data=(X, y))
print(model.summary())

loss, acc = model.evaluate(X, y)

print("Loss: {:.2f}".format(loss))
print("Accuracy: {:.2f}%".format(acc*100))

I get the following error:

ValueError: Input 0 is incompatible with layer lstm_10: expected ndim=3, found ndim=2

The full error traceback:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-35-58fa9218c3f3> in <module>
     31 model.fit(X, y, 
     32           epochs=3,
---> 33           validation_data=(X, y))
     34 print(model.summary())
     35 

C:\Anaconda3\lib\site-packages\keras\engine\training.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, **kwargs)
    950             sample_weight=sample_weight,
    951             class_weight=class_weight,
--> 952             batch_size=batch_size)
    953         # Prepare validation data.
    954         do_validation = False

C:\Anaconda3\lib\site-packages\keras\engine\training.py in _standardize_user_data(self, x, y, sample_weight, class_weight, check_array_lengths, batch_size)
    675             # to match the value shapes.
    676             if not self.inputs:
--> 677                 self._set_inputs(x)
    678 
    679         if y is not None:

C:\Anaconda3\lib\site-packages\keras\engine\training.py in _set_inputs(self, inputs, outputs, training)
    587                 assert len(inputs) == 1
    588                 inputs = inputs[0]
--> 589             self.build(input_shape=(None,) + inputs.shape[1:])
    590             return
    591 

C:\Anaconda3\lib\site-packages\keras\engine\sequential.py in build(self, input_shape)
    219             self.inputs = [x]
    220             for layer in self._layers:
--> 221                 x = layer(x)
    222             self.outputs = [x]
    223             self._build_input_shape = input_shape

C:\Anaconda3\lib\site-packages\keras\layers\recurrent.py in __call__(self, inputs, initial_state, constants, **kwargs)
    530 
    531         if initial_state is None and constants is None:
--> 532             return super(RNN, self).__call__(inputs, **kwargs)
    533 
    534         # If any of `initial_state` or `constants` are specified and are Keras

C:\Anaconda3\lib\site-packages\keras\engine\base_layer.py in __call__(self, inputs, **kwargs)
    412                 # Raise exceptions in case the input is not compatible
    413                 # with the input_spec specified in the layer constructor.
--> 414                 self.assert_input_compatibility(inputs)
    415 
    416                 # Collect input shapes to build layer.

C:\Anaconda3\lib\site-packages\keras\engine\base_layer.py in assert_input_compatibility(self, inputs)
    309                                      self.name + ': expected ndim=' +
    310                                      str(spec.ndim) + ', found ndim=' +
--> 311                                      str(K.ndim(x)))
    312             if spec.max_ndim is not None:
    313                 ndim = K.ndim(x)

ValueError: Input 0 is incompatible with layer lstm_10: expected ndim=3, found ndim=2

It seems that the dimensions of the output of the first LSTM layer (supposedly dim=2) does not fit together with the required input dimensions of the second LSTM layer (dim=3 for batch, timesteps, features).

What bugs me is that adding LSTM layers together the way I did it seems to work here for example: https://adventuresinmachinelearning.com/keras-lstm-tutorial/

The model works when I remove the second LSTM layer.

Answer 1

By default LSTM only returns it's final output after the last element of a sequence. If you want to chain two together then you need to pass the output after each element of the sequence from the first LSTM to the second. eg

model = Sequential()
model.add(LSTM(5, return_sequences=True))
model.add(LSTM(5, activation="softmax"))

See the docs for details on how return_sequence works https://keras.io/layers/recurrent/