5b6da0cf476a8adf7320766760a0756d53df0e48,opennmt/decoders/self_attention_decoder.py,SelfAttentionDecoder,decode,#SelfAttentionDecoder#Any#Any#Any#Any#Any#Any#Any#Any#,37

Before Change

             memory=None,
             memory_sequence_length=None,
             return_logits=True):
    with tf.variable_scope("position_embedding"):
      input_dim = inputs.get_shape().as_list()[-1]
      position_embedding = create_position_embedding(
        input_dim,
        128,
        sequence_length)
      inputs = self.position_encoding_reducer.reduce(inputs, position_embedding)

    inputs = tf.layers.dropout(
      inputs,
      rate=self.dropout,
      training=mode == tf.estimator.ModeKeys.TRAIN)

After Change

    self.dropout = dropout
    self.position_encoder = position_encoder

  def decode(self,
             inputs,
             sequence_length,
             vocab_size,
             encoder_states=None,
             mode=tf.estimator.ModeKeys.TRAIN,
             memory=None,
             memory_sequence_length=None,
             return_logits=True):
    if self.position_encoder is not None:
      inputs = self.position_encoder(inputs, sequence_length=sequence_length)

    inputs = tf.layers.dropout(
      inputs,
      rate=self.dropout,
      training=mode == tf.estimator.ModeKeys.TRAIN)

Instances