def subject_network(self, x):
'''
:param x: query embedding after look-up
:return:subject_score:[batch_size,cand_sub_size] loss_subject:[1]
'''
with tf.variable_scope("subject_part"):
_, query4subject = NN.bi_gru(x, self.x_lens, self.n_hidden,
"bi_gru4subject_query",
self.keep_prob, self.is_training)
query4subject = tf.concat(query4subject, -1)
# [batch,type_len]
query4subject = tf.sigmoid(
NN.linear(query4subject,
self.qa.type_len,
name="query_trans_subject"))
with tf.variable_scope(tf.get_variable_scope(), reuse=True):
self.type_emb_w_martix = tf.get_variable(
"query_trans_subjectkernel")
self.type_emb_b_martix = tf.get_variable(
"query_trans_subjectbias")
# [batch_size,emb_size]
gold_subject = self.subject_emb[:, 0, :]
self.loss_subject = gold_subject * tf.log(
tf.clip_by_value(
query4subject, 1e-10, 1.0)) + (1 - gold_subject) * tf.log(
tf.clip_by_value(1 - query4subject, 1e-10, 1.0))
# [1]
loss_subject = tf.reduce_mean(
-tf.reduce_sum(self.loss_subject, axis=1), axis=0)
with tf.variable_scope("test"):
# [batch,1,cand_sub]
subject_score = tf.matmul(
tf.expand_dims(query4subject, 1),
tf.transpose(self.subject_emb, perm=[0, 2, 1]))
# [batch,cand_sub,1]
subject_score = tf.squeeze(subject_score, 1)
# subject_score = tf.transpose(subject_score, [0, 2, 1])
return subject_score, loss_subject
def matmul_query_relation(self, query_vec, rel_vec, name):
query_vec = NN.linear(query_vec, int(rel_vec.shape[-1]), name=name)
score = tf.matmul(query_vec, tf.transpose(rel_vec))
# [batch,cand_rel]
return score
