search

Sense-X / Co-DETR

[ICCV 2023] DETRs with Collaborative Hybrid Assignments Training
MIT License
1.01k stars 113 forks source link

Instance Segmentation Results #163

Closed Madave94 closed 1 month ago

Madave94 commented 1 month ago

I was trying to run the instance segmentation Co-DETR model pre-trained on Objects365 for LVIS. For this I used this config _projects/configs/co_dino_vit/co_dino_5scale_lsj_vit_large_lvisinstance.py and this weights https://huggingface.co/zongzhuofan/co-detr-vit-large-lvis-instance .

I ran the following command bash tools/dist_test.sh projects/configs/co_dino_vit/co_dino_5scale_lsj_vit_large_lvis_instance.py weights/co-detr-vit-large-coco-instance.pth 2 --eval segm

Somehow, I got results at 41.1 instead of 60.7 on the instance segmentation results. Any idea what could be wrong? I have attached the logging I received from the standard output.

log_co_detr_lsj_vit_large_lvis_instance.txt

TempleX98 commented 1 month ago
results

I can get correct results using my local codebase. I think there is a bug in this repo and I will check it ASAP.

Madave94 commented 1 month ago

Thanks for the quick reply. I also re-checked my installed libraries. My mmcv-full, mmdet, pytorch and python version align exactly with the README.md specification. Only my cuda version is newer at 12.4, however, as far as I know they are backward compatible, so this should not cause issues.

TempleX98 commented 1 month ago

@Madave94 Hi, I have fixed this bug. Please update the codes and try again.

Madave94 commented 1 month ago

Thanks, with the bug fix I was able to reproduce the results.