DetectionTeamUCAS / RetinaNet_Tensorflow_Rotation

Focal Loss for Dense Rotation Object Detection
MIT License
312 stars 93 forks source link

采用配置文件cfgs_res50_dota_v5(即v1 + iou-smooth l1 loss)cls loss始终在1-1.2之间,reg loss 出现nan #71

Open mingmingzhu123 opened 4 years ago

mingmingzhu123 commented 4 years ago

数据集是HRSC2016 采用配置文件cfgs_res50_dota_v1时loss可以正常收敛,当采用cfgs_res50_dota_v5时cls loss始终在1-1.2之间,reg loss 出现nan,tensorboard中gtboxes_h_gpu、gtboxes_r_gpu_正常有框,positive_anchor也是正常有anchor,但是final_detection_gpu_0始终没有检查框,请问是什么问题?非常感谢!

yangxue0827 commented 4 years ago

https://github.com/Thinklab-SJTU/R3Det_Tensorflow/blob/master/libs/losses/losses.py#L235

使用这个函数比较稳定 @mingmingzhu123