losses.py 文件源码

def calculate_loss(self, predictions, labels, weights=None, **unused_params):
    with tf.name_scope("loss_xent"):
      epsilon = 10e-6
      if FLAGS.label_smoothing:
        float_labels = smoothing(labels)
      else:
        float_labels = tf.cast(labels, tf.float32)
      cross_entropy_loss = float_labels * tf.log(predictions + epsilon) + (
          1 - float_labels) * tf.log(1 - predictions + epsilon)
      cross_entropy_loss = tf.negative(cross_entropy_loss)
      if weights is not None:
        print cross_entropy_loss, weights
        weighted_loss = tf.einsum("ij,i->ij", cross_entropy_loss, weights)
        print "create weighted_loss", weighted_loss
        return tf.reduce_mean(tf.reduce_sum(weighted_loss, 1))
      else:
        return tf.reduce_mean(tf.reduce_sum(cross_entropy_loss, 1))