def aggregate(self, v1, v2, num_classes,
v1_lengths=None, v2_lengths=None, use_masking=True, reuse=False):
"""
Aggregate phase.
Args:
v1: tensor with shape (batch_size, time_steps, num_units)
v2: tensor with shape (batch_size, time_steps, num_units)
num_classes: number of output units
v1_lengths: time_steps in v1
v2_lengths: time_steps in v2
use_masking: use masking
reuse: reuse variables
"""
with tf.variable_scope('aggregate', reuse=reuse) as _:
if use_masking:
v1 = mask_3d(sequences=v1, sequence_lengths=v1_lengths, mask_value=0, dimension=1)
v2 = mask_3d(sequences=v2, sequence_lengths=v2_lengths, mask_value=0, dimension=1)
v1_sum, v2_sum = tf.reduce_sum(v1, [1]), tf.reduce_sum(v2, [1])
v1_v2 = tf.concat(axis=1, values=[v1_sum, v2_sum])
transformed_v1_v2 = self._transform_aggregate(v1_v2, reuse=reuse)
logits = tf.contrib.layers.fully_connected(inputs=transformed_v1_v2,
num_outputs=num_classes,
weights_initializer=tf.random_normal_initializer(0.0, 0.01),
biases_initializer=tf.zeros_initializer(),
activation_fn=None)
return logits
def attend(self, sequence1, sequence2,
sequence1_lengths=None, sequence2_lengths=None, use_masking=True, reuse=False):
"""
Attend phase.
Args:
sequence1: tensor with shape (batch_size, time_steps, num_units)
sequence2: tensor with shape (batch_size, time_steps, num_units)
sequence1_lengths: time_steps in sequence1
sequence2_lengths: time_steps in sequence2
use_masking: use masking
reuse: reuse variables
Returns:
two tensors with shape (batch_size, time_steps, num_units)
"""
with tf.variable_scope('attend') as _:
# tensor with shape (batch_size, time_steps, num_units)
transformed_sequence1 = self._transform_attend(sequence1, reuse)
# tensor with shape (batch_size, time_steps, num_units)
transformed_sequence2 = self._transform_attend(sequence2, True)
# tensor with shape (batch_size, time_steps, time_steps)
self.raw_attentions = tf.matmul(transformed_sequence1, tf.transpose(transformed_sequence2, [0, 2, 1]))
masked_raw_attentions = self.raw_attentions
if use_masking:
masked_raw_attentions = mask_3d(sequences=masked_raw_attentions,
sequence_lengths=sequence2_lengths,
mask_value=- np.inf, dimension=2)
self.attention_sentence1 = attention_softmax3d(masked_raw_attentions)
# tensor with shape (batch_size, time_steps, time_steps)
attention_transposed = tf.transpose(self.raw_attentions, [0, 2, 1])
masked_attention_transposed = attention_transposed
if use_masking:
masked_attention_transposed = mask_3d(sequences=masked_attention_transposed,
sequence_lengths=sequence1_lengths,
mask_value=- np.inf, dimension=2)
self.attention_sentence2 = attention_softmax3d(masked_attention_transposed)
# tensors with shape (batch_size, time_steps, num_units)
alpha = tf.matmul(self.attention_sentence2, sequence1, name='alpha')
beta = tf.matmul(self.attention_sentence1, sequence2, name='beta')
return alpha, beta