Implementing Seq2Seq with GRU in Keras

Question

I implanted the ten-minutes LSTM example from the Keras site and adjusted the network to handle word embeddings instead of character ones (from https://blog.keras.io/a-ten-minute-introduction-to-sequence-to-sequence-learning-in-keras.html ). It worked fine.

But now I struggle with using a GRU instead of an LSTM. After adjusting the variables the compilation and training (fit function) worked. But when I try to use the network to test it with custom inputs it throws:

Dimensions must be equal, but are 232 and 256 for 'add' (op: 'Add') with input shapes: [1,?,?,232], [?,256]

The relevant working code for the LSTM is:

encoder_inputs = Input(shape=(None, num_encoder_tokens), name="Encoder_Input")
encoder = LSTM(latent_dim, return_state=True, name="Encoder_LSTM")
encoder_outputs, state_h, state_c = encoder(encoder_inputs)
encoder_states = [state_h, state_c]
decoder_inputs = Input(shape=(None, num_decoder_tokens), name="Decoder_Input")
decoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True, name="Decoder_LSTM")

decoder_outputs, _, _ = decoder_lstm(decoder_inputs,
                                     initial_state=encoder_states)

decoder_dense = Dense(num_decoder_tokens, activation='softmax', name="DecoderOutput")
decoder_outputs = decoder_dense(decoder_outputs)

model = Model([encoder_inputs, decoder_inputs], decoder_outputs)

model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])
model.summary()

result = model.fit([encoder_input_data, decoder_input_data], decoder_target_data,
          batch_size=batch_size,
          epochs=epochs,
          validation_split=0.2)

encoder_model = Model(encoder_inputs, encoder_states)
decoder_state_input_h = Input(shape=(latent_dim,))
decoder_state_input_c = Input(shape=(latent_dim,))
decoder_states_inputs = [decoder_state_input_h, decoder_state_input_c]
decoder_outputs, state_h, state_c = decoder_lstm(
    decoder_inputs, initial_state=decoder_states_inputs)
decoder_states = [state_h, state_c]
decoder_outputs = decoder_dense(decoder_outputs)
decoder_model = Model(
    [decoder_inputs] + decoder_states_inputs,
    [decoder_outputs] + decoder_states)
reverse_target_word_index = dict(
    (i, word) for word, i in target_token_index.items())

The GRU code is:

encoder_inputs = Input(shape=(None, num_encoder_tokens), name="Encoder_Input")
encoder = GRU(latent_dim, return_state=True, name="Encoder_GRU")
_, encoder_state = encoder(encoder_inputs)
decoder_inputs = Input(shape=(None, num_decoder_tokens), name="Decoder_Input")
decoder_gru = GRU(latent_dim, return_sequences=True, return_state=True, name="Decoder_GRU")

decoder_outputs, _ = decoder_gru(decoder_inputs, initial_state=encoder_state)

decoder_dense = Dense(num_decoder_tokens, activation='softmax', name="DecoderOutput")
decoder_outputs = decoder_dense(decoder_outputs)

model = Model([encoder_inputs, decoder_inputs], decoder_outputs)

model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])
model.summary()

result = model.fit([encoder_input_data, decoder_input_data], decoder_target_data,
          batch_size=batch_size,
          epochs=epochs,
          validation_split=0.2)

encoder_model = Model(encoder_inputs, encoder_state)
decoder_states_inputs = Input(shape=(latent_dim,))
decoder_outputs, decoder_states = decoder_gru(
    decoder_inputs, initial_state=decoder_states_inputs)
decoder_outputs = decoder_dense(decoder_outputs)

decoder_model = Model(
    [decoder_inputs] + decoder_states_inputs,
    [decoder_outputs] + decoder_states) # This is where the error appears

reverse_input_word_index = dict(
    (i, word) for word, i in input_token_index.items())
reverse_target_word_index = dict(
    (i, word) for word, i in target_token_index.items())

I marked the occurrence of the error with "# This is where the error appears".

Thanks for any help that you can give and yes I need to try both system to compere their differences with the given data-set.

Answer 1

decoder_states in your LSTM code is a list so you add list to list resulting in a combined list. But in GRU code you have decoder_states as the output of the GRU layer which will have a different type. Not having full code makes debugging harder but try this: [decoder_outputs] + [decoder_states]) # Notice brackets around decoder_states