def _build_graph(self):
with tf.name_scope('inputs'):
sentence1, sentence1_lengths, sentence2, \
sentence2_lengths, labels, is_training = self._build_inputs()
self._word_embedding = self.make_word_embedding()
sentence1_embed, sentence2_embed, sentence1_rnned, sentence2_rnned = \
self._build_rnn_encoder(
sentence1,
sentence2,
sentence1_lengths,
sentence2_lengths
)
with tf.name_scope('attend'):
sentence1_attend, sentence2_attend, att_weights1, att_weights2 = \
decom_ops.attend(
sentence1_rnned, sentence2_rnned,
sentence1_lengths, sentence2_lengths,
is_training=is_training
)
with tf.name_scope('compare'):
compare1, compare2 = decom_ops.compare(sentence1_embed,
sentence2_embed,
sentence1_attend,
sentence2_attend,
is_training=is_training)
compare_dim = self.config['rnn']['state_size'] * 2
num_category = self.config['data']['num_category']
with tf.name_scope('aggregate'):
result = decom_ops.aggregate(
compare1,
compare2,
sentence1_lengths,
sentence2_lengths,
mapper_num_layers=[compare_dim // 2, num_category],
is_training=is_training)
with tf.name_scope('summary_viz'):
attentions = [
(att_weights1, 'attention_weight1', sentence2_lengths),
(att_weights2, 'attention_weight2', sentence1_lengths)
]
for attention_info in attentions:
self._build_attention_viz(*attention_info)
self.loss = self._build_loss(result, labels)
self.inference_probs = tf.nn.softmax(result)
self.inference = tf.argmax(self.inference_probs, axis=-1)
self.train_step, self.train_op = self._build_train_step(self.loss)
self.summary_op = tf.summary.merge_all()
def _build_graph(self):
with tf.name_scope('inputs'):
sentence1, sentence1_lengths, \
sentence2_pos, sentence2_pos_lengths,\
sentence2_neg, sentence2_neg_lengths,\
is_training = self._build_inputs()
self._word_embedding = self.make_word_embedding()
with tf.name_scope('rnn_encode'):
sentence1_embed, sentence2_pos_embed, sentence2_neg_embed, \
sentence1_rnned, sentence2_pos_rnned, sentence2_neg_rnned = \
self._build_rnn_encoder(
sentence1, sentence2_pos, sentence2_neg,
sentence1_lengths, sentence2_pos_lengths, sentence2_neg_lengths)
with tf.name_scope('attend'):
sentence1_pos_attend, sentence2_pos_attend, \
pos_att_weights1, pos_att_weights2 = \
decom_ops.attend(
sentence1_rnned, sentence2_pos_rnned,
sentence1_lengths, sentence2_pos_lengths,
is_training=is_training)
sentence1_neg_attend, sentence2_neg_attend, \
neg_att_weights1, neg_att_weights2 = \
decom_ops.attend(
sentence1_rnned, sentence2_neg_rnned,
sentence1_lengths, sentence2_neg_lengths,
is_training=is_training,
reuse=True)
with tf.name_scope('compare'):
pos_compare1, pos_compare2 = decom_ops.compare(
sentence1_embed,
sentence2_pos_embed,
sentence1_pos_attend,
sentence2_pos_attend,
is_training=is_training)
neg_compare1, neg_compare2 = decom_ops.compare(
sentence1_embed,
sentence2_neg_embed,
sentence1_neg_attend,
sentence2_neg_attend,
is_training=is_training,
reuse=True)
compare_dim = self.config['rnn']['state_size'] + \
self.config['word']['embedding_dim'] // 2
with tf.name_scope('aggregate'):
pos_result = decom_ops.aggregate(
pos_compare1,
pos_compare2,
sentence1_lengths,
sentence2_pos_lengths,
mapper_num_layers=[int(compare_dim * 0.7), 1],
is_training=is_training)
neg_result = decom_ops.aggregate(
neg_compare1,
neg_compare2,
sentence1_lengths,
sentence2_neg_lengths,
mapper_num_layers=[int(compare_dim * 0.7), 1],
is_training=is_training,
reuse=True)
pos_result = tf.squeeze(tf.tanh(pos_result), axis=[-1])
neg_result = tf.squeeze(tf.tanh(neg_result), axis=[-1])
with tf.name_scope('loss'):
self.loss = self._build_loss(pos_result, neg_result)
with tf.name_scope('summary'):
tf.summary.histogram('pos_result', pos_result)
tf.summary.histogram('neg_result', neg_result)
attentions = [
(pos_att_weights1, 'pos_att_weight1', sentence2_pos_lengths),
(pos_att_weights2, 'pos_att_weight2', sentence1_lengths),
(neg_att_weights1, 'neg_att_weight1', sentence2_neg_lengths),
(neg_att_weights2, 'neg_att_weight2', sentence1_lengths)
]
for attention_info in attentions:
self._build_attention_viz(*attention_info)
self.pos_inference = pos_result
self.neg_inference = neg_result
self.train_step, self.train_op = self._build_train_step(self.loss)
self.summary_op = tf.summary.merge_all()