def rnet_matching_layer(layer_size: int,
att_size: int,
par_vecs: tf.Variable,
qu_vecs: tf.Variable,
par_num_words: tf.Variable,
parallel_iterations: int = 64) -> tf.Variable:
with tf.variable_scope('alignment_par_qu') as scope:
with tf.variable_scope('fw/match_rnn_cell/attention'):
fw_cell = MatchRNNCell(GRUCell(layer_size), qu_vecs, att_size)
with tf.variable_scope('bw/match_rnn_cell/attention'):
bw_cell = MatchRNNCell(GRUCell(layer_size), qu_vecs, att_size)
(fw_out, bw_out), (_, _) = bidirectional_dynamic_rnn(
fw_cell,
bw_cell,
inputs=par_vecs,
dtype=tf.float32,
sequence_length=par_num_words,
scope=scope,
parallel_iterations=parallel_iterations,
swap_memory=True)
match_par_qu_out = tf.concat([fw_out, bw_out], axis=2)
return match_par_qu_out
def match_par_qu_layer(self):
with tf.variable_scope('alignment_par_qu') as scope:
rnn_cell = MatchRNNCell(GRUCell(self.conf_layer_size), self.qu_encoded, self.conf_att_size)
outputs, final_state = dynamic_rnn(rnn_cell, self.par_encoded, self.par_num_words,
parallel_iterations=self.conf_rnn_parallelity,
scope=scope, swap_memory=True, dtype=tf.float32)
with tf.variable_scope('encoding'):
outputs, _, _ = bi_gru_layer([self.conf_layer_size], self.apply_dropout(outputs), self.par_num_words,
self.apply_dropout)
return outputs
def rnet_matching_layer_unidirectional(
layer_size: int,
att_size: int,
par_vecs: tf.Variable,
qu_vecs: tf.Variable,
par_num_words: tf.Variable,
parallel_iterations: int = 64) -> tf.Variable:
with tf.variable_scope('alignment_par_qu') as scope:
with tf.variable_scope('fw/match_rnn_cell/attention'):
rnn_cell = MatchRNNCell(GRUCell(layer_size), qu_vecs, att_size)
output, _ = dynamic_rnn(rnn_cell,
inputs=par_vecs,
dtype=tf.float32,
sequence_length=par_num_words,
scope=scope,
parallel_iterations=parallel_iterations,
swap_memory=True)
return output