Closed xuesongnie closed 1 year ago
Hi, @xuesongnie. Thanks for your question. We haven't frozen the parameters of the backbones in downstream tasks. As for COCO detection tasks, we set norm_eval=True
as ResNet but not freeze stage-1. I hope this would help you.
是否需要冻结前几层的参数,还是全部网络放开训练效果好一点,因为看到代码里面没有写关于冻结参数的部分,所以有这个小问题。