If mmrotate support mmdetection 3.1.0 or above?

[kelvinwang139]

What's the feature?

Our system shall integrate mmyolo, mmrotate,mmpose and they are dependent on mmdetection. For mmrotate, the requirements hint the dependent mmdetecton version is <3.1.0? For align the above projects and try to use the latest mmdetection version, I want to confirm that if mmroate 1.0.0rc1 support mmdetection v3.1.0 or above version. Is any limitation for that?

Any other context?

No response

[231055558]

Hello seniors, I'm sorry that I can't provide some solutions for your problem, because I am also troubled by this matter now. At the same time, I would like to ask whether you can reproduce most of the models normally when you try to run the mmrotate1.x version, because I found the loss of nan in all the models other than RTMDet during the reproduction, and it has not been solved yet. The specific situation is as follows.

My virtual environment based on Linux, python3.8.19 torch1. X (also tried torch2. X), mmcv2.0.0, mmdet3.0.0, mmrotate1.0.0 rc1. I hope that if you are free and have relevant practices, you can give me some guidance.

[231055558]

Hello seniors, I'm sorry that I can't provide some solutions for your problem, because I am also troubled by this matter now. At the same time, I would like to ask whether you can reproduce most of the models normally when you try to run the mmrotate1.x version, because I found the loss of nan in all the models other than RTMDet during the reproduction, and it has not been solved yet. The specific situation is as follows.

My virtual environment based on Linux, python3.8.19 torch1. X (also tried torch2. X), mmcv2.0.0, mmdet3.0.0, mmrotate1.0.0 rc1. I hope that if you are free and have relevant practices, you can give me some guidance.

I have solved this problem by using the image splitting program that comes with mmrotate and the loss is no longer nan

[kelvinwang139]

Hello seniors, I'm sorry that I can't provide some solutions for your problem, because I am also troubled by this matter now. At the same time, I would like to ask whether you can reproduce most of the models normally when you try to run the mmrotate1.x version, because I found the loss of nan in all the models other than RTMDet during the reproduction, and it has not been solved yet. The specific situation is as follows.
My virtual environment based on Linux, python3.8.19 torch1. X (also tried torch2. X), mmcv2.0.0, mmdet3.0.0, mmrotate1.0.0 rc1. I hope that if you are free and have relevant practices, you can give me some guidance.

I have solved this problem by using the image splitting program that comes with mmrotate and the loss is no longer nan

Hi， based on your comments "I found the loss of nan in all the models other than RTMDet during the reproduction, and it has not been solved yet.", if it means RTMDet Rotate model is working well under this project?

Thanks for your answer in advance!

[231055558]

前辈们好，很抱歉，我无法为您的问题提供一些解决方案，因为我现在也被这件事困扰着。同时，我想问一下，当你尝试运行mmrotate1.x版本时，你是否可以正常复现大部分模型，因为在复现过程中，我发现除了RTMDet之外的所有模型都丢失了nan，并且还没有解决。具体情况如下。我的虚拟环境基于 Linux，python3.8.19 torch1。X（也尝试了 torch2。X）、mmcv2.0.0、mmdet3.0.0、mmrotate1.0.0 rc1。我希望如果你有空，有相关的实践，可以给我一些指导。

我通过使用mmrotate自带的图像分割程序解决了这个问题，并且损失不再是nan

嗨，根据您的评论“我在复制过程中发现除 RTMDet 以外的所有模型都丢失了 nan，并且尚未解决”，这是否意味着 RTMDet Rotate 模型在这个项目下运行良好？

提前感谢您的回答！

Sorry I just successfully ran the current version of mmrotate 1.x configuration file with mmdet==3.0.0 and mmcv==2.0.0. I didn't test it with mmdet>=3.1.0
My previous run with loss=nan was due to an error in my split data. After I re-split the problem no longer occurs