mzhe

I also want to know, thanks.

I don’t know if my understanding is correct, looking forward to your reply

 I use the same commad but ge't a map=0.459 But I didnot change anything~

> 我试了直接用带白边的的图片训练，效果极其差。 请问你能训练出结果么，我用0.001的学习率训练 Retina net。bbox loss一直nan

I use projects/configs/co_dino/co_dino_5scale_swin_large_16e_o365tococo.py, and it seems if I freeze the backbone and set the checkpoint to False, it will OOM in a 24G A30

I apologize for the delayed response as I am currently focusing on my graduation. Our algorithm incorporates several image augmentations. To ensure a fair comparison, the first row in tab4...

Ok, I will provide the google link in this week. Btw, the data and environment preparation will be updated this weekend.

@tysnow2022 When I deleted some redundant code, I deleted a line of content by mistake. I just fixed it and tested it, you can pull the latest version code. Sorry...

The training works well in my experiment server. Mention now we only support the one gpu one pair images strategy, which has been introduced in the Readme file.

应该是我之前做可视化安装的一个库，直接注释就可以