352ebffaa5bd74a454c3b2b10121b2d959ade2fc,opennmt/models/sequence_tagger.py,SequenceTagger,_build,#SequenceTagger#Any#Any#Any#Any#,49

Before Change

        // predictions must contain tensors with the same batch size
        // so replicate the transition matrix accordingly.
        transition_params = tf.convert_to_tensor(transition_params)
        transition_params = tf.expand_dims(transition_params, axis=0)
        transition_params = tf.tile(transition_params, [tf.shape(logits)[0], 1, 1])

        predictions["logits"] = logits
        predictions["transition_params"] = transition_params
      else:

After Change

          logits,
          transition_params,
          encoder_sequence_length)
        labels = tf.cast(labels, tf.int64)
      else:
        probs = tf.nn.softmax(logits)
        labels = tf.argmax(probs, axis=2)

      labels_vocab_rev = tf.contrib.lookup.index_to_string_table_from_file(
        self.labels_vocabulary_file,
        vocab_size=self.num_labels)

      predictions = {}
      predictions["length"] = encoder_sequence_length
      predictions["labels"] = labels_vocab_rev.lookup(labels)

      return tf.estimator.EstimatorSpec(
        mode,
        predictions=predictions)

Instances