Closed Keiku closed 2 years ago
I recently created my own tiny_coco dataset. I'm trying to run all the configs and benchmark them, but I get some errors.
Maybe some gt_bboxes of some images are illegal.
@Czm369 I am using the COCO dataset as it is. I don't think it's wrong data.
The dataset generated by the above error is the COCO dataset itself.
Did you use some special data augmentation?
@Czm369 This is the default config distributed by mmdetection v2.20.0. Doesn't it cause any problems in your environment?
The model is normally trained for some iterations, so it may exits some wrong data according to my experience. As you said, the model is trained in your own tiny_coco dataset. And I have not met the problem in my environment
@Czm369 I got an error in tiny_coco, so I returned it to the original COCO dataset and verified it, but the error was still there. If you have time, please look at the config and log properly. Thank you for the advice that tends to be a data issue. I check if there is a problem with the dataset. (But I don't know where the problem is with COCO itself.)
@Keiku This is usually related to cuda. Please update to the latest code implementation, it may be solved by #7090.
@Keiku Hello, did you solve this problem? I'm using the v2.22 code and still have this problem
@ZERO-SPACE-X I'm sorry. I'm busy right now, so I haven't been able to investigate. It will take some time because it includes cuda upgrade and so on.
@Keiku Thanks for the reply, I'm looking for other ways
Thanks for your error report and we appreciate it a lot.
Checklist
Describe the bug A clear and concise description of what the bug is.
The default config tood_r50_fpn_1x_coco.py stops with an error in v2.20.0.
Reproduction
Environment
python mmdet/utils/collect_env.py
to collect necessary environment information and paste it here.$PATH
,$LD_LIBRARY_PATH
,$PYTHONPATH
, etc.)Error traceback If applicable, paste the error trackback here.
Bug fix If you have already identified the reason, you can provide the information here. If you are willing to create a PR to fix it, please also leave a comment here and that would be much appreciated!