loss weighted by num_pos

[plutospp]

我在train.py看到这段代码

               for l in range(len(outputs)):
                    loss_item, num_pos = yolo_loss(l, outputs[l], targets)
                    loss_value_all  += loss_item
                    num_pos_all     += num_pos
                loss_value  = loss_value_all / num_pos_all

对正样本做反向加权的用意是? 这是原作使用的trick吗?

[bubbliiiing]

归一化