def contextual_embedding(self):
"""
contextual embedding
:return:
"""
with tf.variable_scope('paragraph_encoding'):
self.h = rnn(self.x_embed, self.hidden_size, self.x_length)
with tf.variable_scope('question_enconding'):
self.u = rnn(self.q_embed, self.hidden_size, self.q_length)
if self.use_dropout:
self.h = tf.nn.dropout(self.h, self.dropout_keep_prob)
self.u = tf.nn.dropout(self.u, self.dropout_keep_prob)
def encode(self):
with tf.variable_scope("passage_encoding"):
self.u_p = rnn('gru',
self.p_embed,
self.hidden_size,
self.p_length,
layer_num=1)
with tf.variable_scope("question_encoding"):
self.u_q = rnn('gru',
self.q_embed,
self.hidden_size,
self.q_length,
layer_num=1)
if self.use_dropout:
self.u_p = tf.nn.dropout(self.u_p, self.dropout_keep_prob)
self.u_q = tf.nn.dropout(self.u_q, self.dropout_keep_prob)
def output(self):
self.p1 = linear(self.hidden_size * 10, self.g, self.m, '1')
with tf.variable_scope("output_rnn"):
m_ = rnn('lstm',
self.m,
self.hidden_size,
self.x_length,
layer_num=1)
self.p2 = linear(self.hidden_size * 10, self.g, m_, '2')
def modeling(self):
with tf.variable_scope("modeling"):
self.m = rnn('lstm',
self.g,
self.hidden_size,
self.x_length,
layer_num=1)
if self.use_dropout:
self.m = tf.nn.dropout(self.m, self.dropout_keep_prob)
def output_layer(self):
params = [([self.params["w_h_p"],
self.params["w_h_a"]], self.params["v"]),
([self.params["w_u_q_"],
self.params["w_v_q"]], self.params["v"])]
with tf.variable_scope("output_layer"):
self.h_p = self.h_p + rnn("gru",
self.h_p,
self.hidden_size,
self.p_length,
dropout_keep_prob=self.dropout_keep_prob)
self.p1, self.p2 = answer_pointer(self.h_p, self.p_length, self.u_q,
self.q_length, self.hidden_size,
params, self.batch_size)