losses.py 文件源码

def contrastive_loss(labels, logits, margin_gen=0, margin_imp=1, scope=None):
    """With this definition the loss will be calculated.
        Args:
          y: The labels.
          distance: The distance vector between the output features..
          batch_size: the batch size is necessary because the loss calculation would be over each batch.
        Returns:
          The total loss.
    """
    with ops.name_scope(scope, "contrastive_loss", [labels, logits]) as scope:
        # logits.get_shape().assert_is_compatible_with(onehot_labels.get_shape())

        labels = math_ops.cast(labels, logits.dtype)

        # term_1 = tf.multiply(labels, tf.square(logits))
        term_1 = tf.multiply(labels, tf.square(tf.maximum((logits - margin_gen), 0)))
        term_2 = tf.multiply(1 - labels, tf.square(tf.maximum((margin_imp - logits), 0)))

        # Contrastive
        Contrastive_Loss = tf.add(term_1, term_2) / 2
        loss = tf.losses.compute_weighted_loss(Contrastive_Loss, scope=scope)

        return loss