expected_classification_loss = (batch_size * num_anchors
                                    * (num_classes+1) * np.log(2.0))