LSTM Encoder-Decoder Inference Model

Question

Many tutorials for seq2seq encoder-decoder architecture based on LSTM, (for example English-French translation), define the model as follow:

encoder_inputs = Input(shape=(None,))
en_x=  Embedding(num_encoder_tokens, embedding_size)(encoder_inputs)

# Encoder lstm
encoder = LSTM(50, return_state=True)
encoder_outputs, state_h, state_c = encoder(en_x)

# We discard `encoder_outputs` and only keep the states.
encoder_states = [state_h, state_c]

# Set up the decoder, using `encoder_states` as initial state.
decoder_inputs = Input(shape=(None,))

# french word embeddings
dex=  Embedding(num_decoder_tokens, embedding_size)
final_dex= dex(decoder_inputs)

# decoder lstm
decoder_lstm = LSTM(50, return_sequences=True, return_state=True)
decoder_outputs, _, _ = decoder_lstm(final_dex,
                                     initial_state=encoder_states)

decoder_dense = Dense(num_decoder_tokens, activation='softmax')
decoder_outputs = decoder_dense(decoder_outputs)

# While training, model takes eng and french words and outputs #translated french word
fullmodel = Model([encoder_inputs, decoder_inputs], decoder_outputs)

# rmsprop is preferred for nlp tasks
fullmodel.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['acc'])

fullmodel.fit([encoder_input_data, decoder_input_data], decoder_target_data,
          batch_size=128,
          epochs=100,
          validation_split=0.20)

Then for prediction, they define infernce models as follow:

# define the encoder model 
encoder_model = Model(encoder_inputs, encoder_states)
encoder_model.summary()


# Redefine the decoder model with decoder will be getting below inputs from encoder while in prediction
decoder_state_input_h = Input(shape=(50,))
decoder_state_input_c = Input(shape=(50,))
decoder_states_inputs = [decoder_state_input_h, decoder_state_input_c]
final_dex2= dex(decoder_inputs)

decoder_outputs2, state_h2, state_c2 = decoder_lstm(final_dex2, initial_state=decoder_states_inputs)

decoder_states2 = [state_h2, state_c2]
decoder_outputs2 = decoder_dense(decoder_outputs2)

# sampling model will take encoder states and decoder_input(seed initially) and output the predictions(french word index) We dont care about decoder_states2
decoder_model = Model(
    [decoder_inputs] + decoder_states_inputs,
    [decoder_outputs2] + decoder_states2)

Then predict using:

# Reverse-lookup token index to decode sequences back to
# something readable.
reverse_input_char_index = dict(
    (i, char) for char, i in input_token_index.items())
reverse_target_char_index = dict(
    (i, char) for char, i in target_token_index.items())


def decode_sequence(input_seq):
    # Encode the input as state vectors.
    states_value = encoder_model.predict(input_seq)

    # Generate empty target sequence of length 1.
    target_seq = np.zeros((1,1))
    # Populate the first character of target sequence with the start character.
    target_seq[0, 0] = target_token_index['START_']

   # Sampling loop for a batch of sequences
    # (to simplify, here we assume a batch of size 1).
    stop_condition = False
    decoded_sentence = ''
    while not stop_condition:
        output_tokens, h, c = decoder_model.predict(
            [target_seq] + states_value)

        # Sample a token
        sampled_token_index = np.argmax(output_tokens[0, -1, :])
        sampled_char = reverse_target_char_index[sampled_token_index]
        decoded_sentence += ' '+sampled_char

        # Exit condition: either hit max length
        # or find stop character.
        if (sampled_char == '_END' or
           len(decoded_sentence) > 52):
            stop_condition = True

       # Update the target sequence (of length 1).
        target_seq = np.zeros((1,1))
        target_seq[0, 0] = sampled_token_index

        # Update states
        states_value = [h, c]
return decoded_sentence

My question is, they trained the model with the name 'fullmodel' to get best weights ... in prediction part, they used the inference models with names (encoder_model & decoder_model) ... so they didn't use any weights from the 'fullmodel' ?!

I don't understand how they benefit from the trained model!

Answer 1

If you notice carefully, the trained layer weights are being reused. For example, while creating decoder_model we use decoder_lstm layer which was defined as a part of full model, decoder_outputs2, state_h2, state_c2 = decoder_lstm(final_dex2, initial_state=decoder_states_inputs),

and encoder model too uses, encoder_inputs and encoder_states layer defined previously. encoder_model = Model(encoder_inputs, encoder_states)

Due to the architecture of the encoder-decoder model, we need to perform these implementations hacks. Also, as the keras documentation mentions, With the functional API, it is easy to reuse trained models: you can treat any model as if it were a layer, by calling it on a tensor. Note that by calling a model you aren't just reusing the architecture of the model, you are also reusing its weights. For more details refer - https://keras.io/getting-started/functional-api-guide/#all-models-are-callable-just-like-layers

Answer 2

The trick is that everything is in the same variable scope, so the variables got reused.