Closed jinweiLiu closed 3 years ago
Hello! When I start training, the following issue occurred “FloatingPointError: Predicted boxes or scores contain Inf/NaN. Training has diverged.” What should I do?
Hello! Have you solved your problem?
decrease the learning rate may help
decrease the learning rate may help
I've tried this method, but it didn't solve my problem. Although, thank you for your reply.
Hi,
sorry for replying to the thread late. Could you let me know what your setup is? Like your Pytorch, detectron2, and cudnn version? The code was tested in Pytorch 1.6 and Detectron2 v0.2.1. Maybe you can check whether the problem can be resolved if an older version is used?
Also, if you have any additional information to share to resolve the issue, please let us know. We will look into the issue on our end as well.
Thanks, Xin
On Tue, Apr 13, 2021 at 8:36 PM jgsimard @.***> wrote:
@diaosilei https://github.com/diaosilei , @jinweiLiu https://github.com/jinweiLiu did you solve the problem ? I get the the same thing! The training starts but it diverges, even if I decrease the learning rate
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/ucbdrive/few-shot-object-detection/issues/104#issuecomment-819204391, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABVXTBADMJNWSOSJOUIL223TIUESVANCNFSM42QKUQRA .
you could try to set the learning rate lower发自我的荣耀手机-------- 原始邮件 --------发件人: jgsimard @.>日期: 2021年4月14日周三 11:36收件人: ucbdrive/few-shot-object-detection @.>抄送: jinweiLiu @.>, Mention @.>主 题: Re: [ucbdrive/few-shot-object-detection] FloatingPointError (#104) @diaosilei , @jinweiLiu did you solve the problem ? I get the the same thing! The training starts but it diverges, even if I decrease the learning rate
—You are receiving this because you were mentioned.Reply to this email directly, view it on GitHub, or unsubscribe.
Hi, I think it it a mismatch with detectron v.04. It works well with detectron v0.3 !
Great news! @thomasehuang and I would also keep an eye on updating the codebase to the latest detectron2. It's a bit unfortunate that sometimes these library versions are not completely backward compatible.
Hello! When I start training, the following issue occurred “FloatingPointError: Predicted boxes or scores contain Inf/NaN. Training has diverged.” What should I do?