loss不下降

[mandy-w]

我最近用cascade-fpn训练自己的一个数据集，训练这个数据集的时候我把学习率已经调到了0.00001，loss一直在0.4左右不能下降，不知道是什么原因呢？cascade是不是有一定的数据集使用范围？有些数据集并不适用呢·？

[zhaoweicai]

How about mAP? Cascade R-CNN has been widely tested in many different datasets by different people. It should be very robust. However, it is still possible that it doesn't work in some cases.

[mandy-w]

是一个检测刀的数据集，这个数据集我在faster-rcnn上训练能达到mAP=0.83左右，但是cascade-rcnn上mAP只有0.78左右

[zhaoweicai]

Cascade R-CNN should not be worse than faster-rcnn, since you can set the weights of the 2nd/3rd stages to 0. You may need to tune the hyperparameters a little bit, e.g. IoU thresholds. Or maybe try 2-stage cascade first.

[mandy-w]

非常感谢你耐心的解答，我再试试调一下