MLSTM to match passage encodes with question encodes

mlstm.py

# Python 3.6 + TF 1.6
# Han Xiao ([email protected])

import tensorflow as tf
import tensorflow.contrib as tc


#### Usage: matching passage encodes with question encodes
ml = MatchLSTMLayer(hidden_size=16,
                    control_gate=False,
                    pooling_window=5,  # do pooling otherwise attention on long passage will give OOM
                    name='demo',
                    act_fn=tf.nn.relu,
                    attend_hidden_size=16)

ml.match(input_encodes=p_encodes,
         attended_encodes=q_encodes,
         input_length=p_length,
         input_mask=p_mask,
         attended_mask=q_mask)


#### Details

def attend_pooling(pooling_vectors, ref_vector, hidden_size, scope=None,
                   pooling_mask=None, activation_fn=tf.tanh, output_logit=False):
    """
    Applies attend pooling to a set of vectors according to a reference vector.
    Args:
        pooling_vectors: the vectors to pool in B x T x D size
        ref_vector: the reference vector in B x D size, at a single time t, D can be different than pooling_vectors
        hidden_size: the hidden size for attention function
        scope: score name
    Returns:
        the pooled vector in B x D size
    """
    with tf.variable_scope(scope or 'attend_pooling'):
        # pooling vectors must be B x T x D size
        assert pooling_vectors.get_shape().ndims == 3
        # ref_vector must be B x D size, at a single time t
        assert ref_vector.get_shape().ndims == 2
        if pooling_mask is not None:
            # pooling_mask must be B x T size
            assert pooling_mask.get_shape().ndims == 2
        U = activation_fn(tc.layers.fully_connected(pooling_vectors,
                                                    num_outputs=hidden_size,
                                                    activation_fn=None)
                          + tf.expand_dims(tc.layers.fully_connected(ref_vector,
                                                                     num_outputs=hidden_size,
                                                                     activation_fn=None), axis=1))
        logits = tc.layers.fully_connected(U, num_outputs=1, activation_fn=None)
        if pooling_mask is not None:
            logits -= tf.expand_dims(1.0 - pooling_mask, axis=2) * 1e30
        scores = tf.nn.softmax(logits, 1)

        pooled_vector = tf.reduce_sum(pooling_vectors * scores, axis=1)

        # pooled vector is B x D size, score is B x L size
    return pooled_vector, logits if output_logit else scores


class MatchLSTMAttnCell(tc.rnn.LSTMCell):
    """
    Implements the Match-LSTM attention cell
    """

    def __init__(self, num_units, context_to_attend, control_gate, pooling_window, attend_mask, act_fn,
                 attend_hidden_size):
        super().__init__(num_units, state_is_tuple=True)
        if pooling_window:
            self.context_to_attend = tf.nn.pool(context_to_attend,
                                                window_shape=[pooling_window],
                                                strides=[pooling_window],
                                                pooling_type='MAX', padding='SAME',
                                                name='max_pool_on_context')
            # do the same max-pooling for mask
            # mask is usually BxT. To do pooling, we need to first expand it to BxTx1
            # then squeeze it back to BxT
            self.attend_mask = tf.squeeze(tf.nn.pool(tf.expand_dims(attend_mask, -1),
                                                     window_shape=[pooling_window],
                                                     strides=[pooling_window],
                                                     pooling_type='MAX', padding='SAME',
                                                     name='max_pool_on_context_mask'), -1)
        else:
            self.context_to_attend = context_to_attend
            self.attend_mask = attend_mask

        self.control_gate = control_gate
        self.act_fn = act_fn
        self.attend_hidden_size = attend_hidden_size

    def __call__(self, inputs, state, scope=None):
        (c_prev, h_prev) = state
        with tf.variable_scope(scope or type(self).__name__):
            ref_vector = tf.concat([inputs, h_prev], -1)
            attended_context, scores = attend_pooling(self.context_to_attend,
                                                      ref_vector,
                                                      self.attend_hidden_size,
                                                      pooling_mask=self.attend_mask,
                                                      activation_fn=self.act_fn)

            new_inputs = tf.concat([inputs, attended_context,
                                    inputs - attended_context,
                                    inputs * attended_context],
                                   -1)

            if self.control_gate:
                # modified by adding another gate to the input
                control_gate = tc.layers.fully_connected(new_inputs,
                                                         num_outputs=self.output_size * 8,
                                                         activation_fn=tf.nn.sigmoid)
                new_inputs *= control_gate

            return super().__call__(new_inputs, state, scope)


class MatchLSTMLayer:
    """
    Implements the Match-LSTM layer, which attend to the question dynamically in a LSTM fashion.
    """

    def __init__(self, hidden_size: int, control_gate: bool,
                 pooling_window: int, name: str, act_fn, attend_hidden_size: int):
        self.hidden_size = hidden_size
        self.output_size = hidden_size * 2  # bi-directional
        self.act_fn = act_fn
        self.name = name
        self.control_gate = control_gate
        self.pooling_window = pooling_window  # useful when context is too long.
        self.attend_hidden_size = attend_hidden_size

    def match(self, input_encodes, attended_encodes, input_length, input_mask, attended_mask):
        """
        Match the passage_encodes with question_encodes using Match-LSTM algorithm
        """
        with tf.variable_scope(self.name):
            cell_fw = MatchLSTMAttnCell(self.hidden_size, attended_encodes, self.control_gate, self.pooling_window,
                                        attended_mask, self.act_fn, self.attend_hidden_size)
            cell_bw = MatchLSTMAttnCell(self.hidden_size, attended_encodes, self.control_gate, self.pooling_window,
                                        attended_mask, self.act_fn, self.attend_hidden_size)
            outputs, state = tf.nn.bidirectional_dynamic_rnn(cell_fw, cell_bw,
                                                             inputs=input_encodes,
                                                             sequence_length=input_length,
                                                             dtype=tf.float32)
            match_outputs = tf.concat(outputs, 2)
            state_fw, state_bw = state
            c_fw, h_fw = state_fw
            c_bw, h_bw = state_bw
            match_state = tf.concat([h_fw, h_bw], 1)
        return match_outputs, match_state