loss_ id is Nan

[yangwhf]

loss id and total loss sometimes becomes Nan. I'm not sure if this is normal

[HanGuangXin]

Provide more information, please.

---Original--- From: @.> Date: Tue, Jun 14, 2022 11:09 AM To: @.>; Cc: @.***>; Subject: [HanGuangXin/ByteTrackReID] loss id is Nan (Issue #25)

loss id and total loss sometimes becomes Nan. I'm not sure if this is normal

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>

[yangwhf]

请提供更多信息。 … ---原创---来自：@.> 日期：2022 年 6 月 14 日星期二上午 11:09 至：@.>; 抄送：@.>; 主题：[汉光新/ByteTrack_ReID] loss_id is Nan（问题#25）loss_id和total_loss有时会变成Nan。我不确定这是否正常 — 直接回复此电子邮件、在 GitHub 上查看或取消订阅。您收到此消息是因为您订阅了此线程。消息 ID：@.>

Use "MOT17 half train" and "CrowdHuman" mixed dataset training, use "exps/example/mot/yolox_x_ablation.py" and modified according to "exps/example/mot/yolox_s_mot17_half.py"

[HanGuangXin]

It is normal when you training with mixed MOT datasets and detection datasets.

If you do not want it happen, maybe try larger batchsize.

[yangwhf]

当您使用混合的 MOT 数据集和检测数据集进行训练时，这是正常的。

如果您不希望它发生，可以尝试更大的批量。

Thanks for your answer, I'm confused why this is happening?

[HanGuangXin]

Detection datasets do not has ID annotations. That's all.

[yangwhf]

检测数据集没有 ID 注释。就这样。

I see, thank you for your answer

[HanGuangXin]

:-)