linxiaocong
linxiaocong
> 因为太多的反馈训练问题,为了避免我的代码对大家的误导,抄作业了训练和loss,一直抄到 label smooth 之前。后面这个repo不会再抄和更新,抄到这里是为了发现和定位问题。 > **为了模型有更专业的持续维护,请大家用原作者的完善版本。** > > qqwwee版的训练确实很多坑,抄作业过程发现的问题和修改: > (1) 模型主干没有太大问题,修改了kernel_initializer和 L2正则化(会加速收敛,酌情加上可能更好); > (2) train里面的letterbox预处理,不如scale方式准确; > (3) 没有多尺度输入; > (4) 之前的ciou loss也不太对,miemie2013的全尺度 ciou loss 好很多; > (5) 从主干输出解码的流程也有精度问题;...
> > > Hey guys, > > > Sorry for the late reply. I was busy with some things. [Here](https://drive.google.com/drive/folders/1bkyaaXbfpm_c8Krfr3i2SNMA9VVcCu8W?usp=sharing) is my checkpoint I got the results from. I have...
I find: 1、in train.py and test_wholeimage_swapsingle.py, the initial network was not the same... in train.py: ``` from models.projected_model import fsModel # model initial model = fsModel() # fsModel() source from...
> so glad, can you provide your email? or another way?