common_layers.py 文件源码

def weights_concatenated(labels):
  """Assign weight 1.0 to the "target" part of the concatenated labels.

  The labels look like:
    source English I love you . ID1 target French Je t'aime . ID1 source
      English the cat ID1 target French le chat ID1 source English ...

  We want to assign weight 1.0 to all words in the target text (including the
  ID1 end symbol), but not to the source text or the boilerplate.  In the
  above example, the target words that get positive weight are:
    Je t'aime . ID1 le chat ID1

  Args:
    labels: a Tensor
  Returns:
    a Tensor
  """
  eos_mask = tf.to_int32(tf.equal(labels, 1))
  sentence_num = tf.cumsum(eos_mask, axis=1, exclusive=True)
  in_target = tf.equal(tf.mod(sentence_num, 2), 1)
  # first two tokens of each sentence are boilerplate.
  sentence_num_plus_one = sentence_num + 1
  shifted = tf.pad(sentence_num_plus_one,
                   [[0, 0], [2, 0], [0, 0], [0, 0]])[:, :-2, :, :]
  nonboilerplate = tf.equal(sentence_num_plus_one, shifted)
  ret = tf.to_float(tf.logical_and(nonboilerplate, in_target))
  return ret