Object detection Task with OBB: 前期训练过程正常,Loss正常下降,奇怪的是会在某一个batch 的时间节点,Box Loss突变为inf,然后在这个batch之后loss全变为NaN,甚至直接影响到分类Loss,问题排查过程如下: 1. 考虑到类别不均衡问题,梯度易爆炸,降低学习率与batch size 0.01每次减小一个数量级->0.00001,最后调整为0 ...
Search before asking I have searched the YOLOv5 issues and found no similar bug report. YOLOv5 Component Training, Validation Bug The problem occurs when the prediction has 0.0 values. These then follow through the method bbox_iou in NAN...