回归损失 loss=inf

[lyfighting]

您好！感谢您的工作。我将该网络在coco2017数据集上训练人车检测模型，将anchor数量由1个变为6个，训练中在验证集上出现回归损失loss=inf的情况，而分类损失数值是正常的。请问可能是什么地方出了问题？ 谢谢解答！

[scutzhe]

I met same problem, and I don't know how to solve. have you solved?

[hufangjian]

学习率调小一点。

[torta24x]

As of now I have found two reasons for Inf loss value either during train or validation

1. Higher learning rate : try reducing learning rate , say do 1/10th or 1/100th etc

2. Issue with input data : This is most common in custom dataset Usually the size of face bounding box is too small . In my case the bounding box width and height where zero (0) leading loss to be inf.

Hope that helps