def _encode(self):
with tf.variable_scope('encoding', reuse=tf.AUTO_REUSE):
self.seq_encode, _ = cu_rnn('bi-lstm', self.token_emb,
self.hidden_size, self.batch_size,
self.layer_num)
if self.is_train:
self.seq_encode = tf.nn.dropout(self.seq_encode,
self.dropout_keep_prob)
def _encoder(self):
with tf.variable_scope('encoder'):
if self.args.encoder_type == 'rnn':
y, _ = cu_rnn('bi-gru', self.token_emb, int(self.args.n_emb / 2), self.n_batch, self.args.n_layer)
elif self.args.encoder_type == 'cnn':
y = cnn(self.token_emb, self.mask, self.args.n_emb, 3)
elif self.args.encoder_type == 'ffn':
y = ffn(self.token_emb, int(self.args.n_emb * 2), self.args.n_emb,
self.args.dropout_keep_prob if self.is_train else 1)
self.token_encoder = residual_link(self.token_emb, y, self.args.dropout_keep_prob if self.is_train else 1.0)
def _encode(self):
with tf.variable_scope('encoding', reuse=tf.AUTO_REUSE):
self.H, _ = cu_rnn('bi-lstm', self.token_emb, self.n_hidden,
self.n_batch, self.n_layer)
if self.is_train:
self.H = tf.nn.dropout(self.H, rate=1 - self.dropout_keep_prob)