search

megvii-research / PETR

[ECCV2022] PETR: Position Embedding Transformation for Multi-View 3D Object Detection & [ICCV2023] PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images
Other
857 stars 133 forks source link

grad_norm:nan #141

Open losetiger opened 1 year ago

losetiger commented 1 year ago

我在petr_r50dcn_gridmask_p4.py 将img_backbone换成了resnext/darknet 但是不管我把学习率调的多小 训练过程中打印的grad_norm都是nan是怎么回事呢 有没有什么解决办法呢

sidiangongyuan commented 9 months ago

请问解决了么

zhao913 commented 5 months ago

Hello, I recently had the same problem, did you solve it?

losetiger commented 5 months ago

你可以尝试调试一下梯度裁剪的参数 但是 虽然这个问题解决了 训练出来的模型准确率一直为零。。。。 我也不知道啥情况 我是尝试修改backbone的 

心ф冇ba砍伱の刀 @.***

 

------------------ 原始邮件 ------------------ 发件人: @.***>; 发送时间: 2024年4月17日(星期三) 上午9:00 收件人: "megvii-research/PETR"; 抄送: "Author"; 主题: Re: [megvii-research/PETR] grad_norm:nan (Issue #141)

Hello, I recently had the same problem, did you solve it?

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