How to use AttentionMechanism with MultiRNNCell and dynamic_decode?

Question

I want to create a multi-layered dynamic RNN-based decoder that uses an attention mechanism. To do this, I first create an attention mechanism:

attention_mechanism = BahdanauAttention(num_units=ATTENTION_UNITS,
                                        memory=encoder_outputs,
                                        normalize=True)

Then I use the AttentionWrapper to wrap a LSTM cell with the attention mechanism:

attention_wrapper = AttentionWrapper(cell=self._create_lstm_cell(DECODER_SIZE),
                                             attention_mechanism=attention_mechanism,
                                             output_attention=False,
                                             alignment_history=True,
                                             attention_layer_size=ATTENTION_LAYER_SIZE)

where self._create_lstm_cell is defined as follows:

@staticmethod
def _create_lstm_cell(cell_size):
    return BasicLSTMCell(cell_size)

I then do some bookkeeping (e.g. creating my MultiRNNCell, creating an initial state, creating a TrainingHelper, etc.)

        attention_zero = attention_wrapper.zero_state(batch_size=tf.flags.FLAGS.batch_size, dtype=tf.float32)

        # define initial state
        initial_state = attention_zero.clone(cell_state=encoder_final_states[0])

        training_helper = TrainingHelper(inputs=self.y,  # feed in ground truth
                                         sequence_length=self.y_lengths)  # feed in sequence lengths

        layered_cell = MultiRNNCell(
            [attention_wrapper] + [ResidualWrapper(self._create_lstm_cell(cell_size=DECODER_SIZE))
                                   for _ in range(NUMBER_OF_DECODER_LAYERS - 1)])

        decoder = BasicDecoder(cell=layered_cell,
                               helper=training_helper,
                               initial_state=initial_state)

        decoder_outputs, decoder_final_state, decoder_final_sequence_lengths = dynamic_decode(decoder=decoder,
                                                                                              maximum_iterations=tf.flags.FLAGS.max_number_of_scans // 12,
                                                                                              impute_finished=True)

But I receive the following error: AttributeError: 'LSTMStateTuple' object has no attribute 'attention'.

What is the correct way to add an attention mechanism to a MultiRNNCell dynamic decoder?

Answer 1

Have you tried using the attention wrapper provided by tf.contrib?

Here is an example using both an attention wrapper and dropout:

cells = []
for i in range(n_layers):                   
    cell = tf.contrib.rnn.LSTMCell(n_hidden, state_is_tuple=True)

    cell = tf.contrib.rnn.AttentionCellWrapper(
        cell, attn_length=40, state_is_tuple=True)

    cell = tf.contrib.rnn.DropoutWrapper(cell,output_keep_prob=0.5)
    cells.append(cell)

cell = tf.contrib.rnn.MultiRNNCell(cells, state_is_tuple=True)
init_state = cell.zero_state(batch_size, tf.float32)

Answer 2

What you need to do is that you create the MultiLayer cell then you wrap it with an AttentionWrapper , below is an example :

def decoding_layer(dec_input, encoder_state,
               target_sequence_length, max_target_sequence_length,
               rnn_size,
               num_layers, target_vocab_to_int, target_vocab_size,
               batch_size, keep_prob, decoding_embedding_size , encoder_outputs):
"""
Create decoding layer
:param dec_input: Decoder input
:param encoder_state: Encoder state
:param target_sequence_length: The lengths of each sequence in the target batch
:param max_target_sequence_length: Maximum length of target sequences
:param rnn_size: RNN Size
:param num_layers: Number of layers
:param target_vocab_to_int: Dictionary to go from the target words to an id
:param target_vocab_size: Size of target vocabulary
:param batch_size: The size of the batch
:param keep_prob: Dropout keep probability
:param decoding_embedding_size: Decoding embedding size
:return: Tuple of (Training BasicDecoderOutput, Inference BasicDecoderOutput)
"""
# 1. Decoder Embedding
dec_embeddings = tf.Variable(tf.random_uniform([target_vocab_size, decoding_embedding_size]))
dec_embed_input = tf.nn.embedding_lookup(dec_embeddings, dec_input)

# 2. Construct the decoder cell
def create_cell(rnn_size):
    lstm_cell = tf.contrib.rnn.LSTMCell(rnn_size,
                                        initializer=tf.random_uniform_initializer(-0.1,0.1,seed=2))
    drop = tf.contrib.rnn.DropoutWrapper(lstm_cell, output_keep_prob=keep_prob)
    return drop


dec_cell = tf.contrib.rnn.MultiRNNCell([create_cell(rnn_size) for _ in range(num_layers)])
#dec_cell = tf.contrib.rnn.MultiRNNCell(cells_a)  

#attention details 
attention_mechanism = tf.contrib.seq2seq.BahdanauAttention(num_units=rnn_size, memory=encoder_outputs) 

attn_cell = tf.contrib.seq2seq.AttentionWrapper(dec_cell, attention_mechanism , attention_layer_size=rnn_size/2)

attn_zero = attn_cell.zero_state(batch_size , tf.float32 )

attn_zero = attn_zero.clone(cell_state = encoder_state)

#new_state = tf.contrib.seq2seq.AttentionWrapperState(cell_state = encoder_state, attention = attn_zero  , time = 0 ,alignments=None , alignment_history=())

"""out_cell = tf.contrib.rnn.OutputProjectionWrapper(
            attn_cell, target_vocab_size, reuse=True
        )"""
#end of attention 
#tensor_util.make_tensor_proto(attn_cell)
output_layer = Dense(target_vocab_size,
                     kernel_initializer = tf.truncated_normal_initializer(mean = 0.0, stddev=0.1))

with tf.variable_scope("decode"):
    train_decoder_out = decoding_layer_train(attn_zero, attn_cell, dec_embed_input, 
                     target_sequence_length, max_target_sequence_length, output_layer, keep_prob)

with tf.variable_scope("decode", reuse=True):
    infer_decoder_out = decoding_layer_infer(attn_zero, attn_cell, dec_embeddings, 
                         target_vocab_to_int['<GO>'], target_vocab_to_int['<EOS>'], max_target_sequence_length, 
                         target_vocab_size, output_layer, batch_size, keep_prob)

return (train_decoder_out, infer_decoder_out)