faster-rcnn-pytorch icon indicating copy to clipboard operation
faster-rcnn-pytorch copied to clipboard

ResNet训练mAP问题

Open ChristmasLee opened this issue 2 years ago • 5 comments

b导你好,我在使用Faster-RCNN训练时,分别使用vgg和resnet作为主干网络预训练权重,vgg是正常的,但是resnet在前50个冻结训练正常结束之后,loss变成nan,然后每10个epoch的验证集mAP也降成了0,这是为什么呢? image image

ChristmasLee avatar Jul 30 '22 02:07 ChristmasLee

练飞了……你学习率自己改过吗?

bubbliiiing avatar Jul 31 '22 14:07 bubbliiiing

练飞了……你学习率自己改过吗?

谢谢b导回复,原fasterrcnn您给的是adam优化器: Init_lr = 1e-4 Min_lr = Init_lr * 0.01 optimizer_type = "adam" momentum = 0.9 weight_decay = 0

然后我修改为sgd优化器: Init_lr = 1e-2 Min_lr = Init_lr * 0.01 optimizer_type = "sgd" momentum = 0.937 weight_decay = 5e-4 是按照您给的注释改的呀

ChristmasLee avatar Jul 31 '22 14:07 ChristmasLee

建议别改……

bubbliiiing avatar Jul 31 '22 14:07 bubbliiiing

建议别改……

那b导,您的意思就是fasterrcnn建议只能用adam优化器,不建议sgd是吗

ChristmasLee avatar Jul 31 '22 14:07 ChristmasLee

我建议adam,sgd不好调参

bubbliiiing avatar Aug 01 '22 14:08 bubbliiiing