tf.nn.compute_average_loss