Hi, we have recently updated a new VisDrone training config (gradient clip added), which can avoid NaN now.

[shaoyandea]

          Hi, we have recently updated a new VisDrone training config (gradient clip added), which can avoid NaN now.

Originally posted by @ChenhongyiYang in #40 (comment)
I also meet this problem ,have you solve it?

[ChenhongyiYang]

Hi, may I ask what is the total batch size in your training?

[shaoyandea]

I trained with coco datadet, the total batch size  is 12, and I have solved Nan by adding gradient clip set as visdrone_train.yaml, also the reproduced result is close to result in paper. there is a bug that learning rate always keep original setting in whole training, and iters in reume training is wrong.

…

------------------ 原始邮件 ------------------ 发件人: "ChenhongyiYang/QueryDet-PyTorch" ***@***.***>; 发送时间: 2024年6月4日(星期二) 晚上11:00 ***@***.***>; ***@***.******@***.***>; 主题: Re: [ChenhongyiYang/QueryDet-PyTorch] Hi, we have recently updated a new VisDrone training config (gradient clip added), which can avoid NaN now. (Issue #97) Hi, may I ask what is the total batch size in your training? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>