pem loss nan？

[leemengxing]

thks for your work. but I found that sometimes the batchsize is 1, pem = loss occurs. why?

[leemengxing]

There is a statistical ratio in the loss function. It may happen that the number of iou> 0.7 or iou> 0.9 is 0 for small batch sizes, such as 1, which leads to nan. Is this correct？

[vhvkhoa]

I have also met this problem and have to replace the loss by Focal Loss, I am still on the experiment to make sure it works

[leemengxing]

Have you tried changing the two classifications to multiple classifications for detection tasks instead of proposal tasks? @vhvkhoa

[JJBOY]

There is a statistical ratio in the loss function. It may happen that the number of iou> 0.7 or iou> 0.9 is 0 for small batch sizes, such as 1, which leads to nan. Is this correct？

Yes, you are right.

[lyx190]

I have met this problem when batchsize was set to 16, anyone of you have some ideas to deal with it?

[longchao1]

I have met this problem when batchsize was set to 4, anyone of you have some ideas to deal with it?

[JJBOY]

@lyx190 @longchao1
The reason why NAN loss is raised is explained by @leemengxing. There are two simple ways to solve this problem.
1、Set larger batch size
2、When the batch does not has any positive sample because of the threshold in loss function, just jump this batch

[lyx190]

@lyx190 @longchao1
The reason why NAN loss is raised is explained by @leemengxing. There are two simple ways to solve this problem.
1、Set larger batch size
2、When the batch does not has any positive sample because of the threshold in loss function, just jump this batch

Thanks a lot