-
Notifications
You must be signed in to change notification settings - Fork 427
New issue
Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? # to your account
yolo-wolrd-l 在 coco上finetune 无法成功复现 #76
Comments
Hello, do you use mask-refine when fine-tuning on coco? If I don't use mask-refine, I can't get normal results. Have you ever encountered similar problems? |
@Hudaodao99 @HGao-cv 我这边重新检查一下finetune的config |
@Hudaodao99 你好,我也遇到了同样的问题。同配置coco上微调,达不到论文中的精度,两种微调方式的mAP50比论文低4.2、3.7。 |
你好,我刚尝试用 configs/finetune_coco/yolo_world_l_efficient_neck_2e-4_80e_8gpus_mask-refine_finetune_coco.py的config进行finetune,达到了论文指标。但原文件的权重文件是没有加入cc3m的pretain(这个结果在论文中没有提到),仍旧可以达到map=53.3,这是我疑惑的点。 |
Yes, if I use mask-refine when fine-tuning on coco, I can get the same results as on github. if not, map = 0.486, it's lower. |
@Sally-lxy @Hudaodao99 目前问题看下来应该是 |
Hi all (@Sally-lxy, @Hudaodao99), we have explored the errors about pre-training without You can find more details in configs/finetune_coco, especially for the version without |
您好! 我配置了环境想复现yolo_world_l在coco上的finetune结果,但实际跑出来的结果介于s和m模型结果的中间。
过程详细按照finetune的文档进行实验:
load_from='pretrained_models/yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_train_pretrained-0e566235.pth'
改为load_from='pretrained_models/yolo_world_l_clip_base_dual_vlpan_2e-3adamw_32xb16_100e_o365_goldg_cc3mlite_train_pretrained-7a5eea3b.pth'
02/23 21:54:41 - mmengine - INFO - Epoch(train) [80][800/925] base_lr: 2.0000e-04 lr: 6.9500e-06 eta: 0:01:05 time: 0.5054 data_time: 0.0021 memory: 18624 grad_norm: 886.4797 loss: 350.3436 loss_cls: 97.6750 loss_bbox: 112.2129 loss_dfl: 140.4556
02/23 21:55:06 - mmengine - INFO - Epoch(train) [80][850/925] base_lr: 2.0000e-04 lr: 6.9500e-06 eta: 0:00:39 time: 0.4935 data_time: 0.0021 memory: 18717 grad_norm: 887.4158 loss: 365.0416 loss_cls: 104.4143 loss_bbox: 117.6048 loss_dfl: 143.0225
Corrupt JPEG data: premature end of data segment
02/23 21:55:31 - mmengine - INFO - Epoch(train) [80][900/925] base_lr: 2.0000e-04 lr: 6.9500e-06 eta: 0:00:13 time: 0.5056 data_time: 0.0021 memory: 18637 grad_norm: 881.9796 loss: 357.7269 loss_cls: 102.4538 loss_bbox: 112.8057 loss_dfl: 142.4673
02/23 21:57:06 - mmengine - INFO - bbox_mAP_copypaste: 0.486 0.656 0.530 0.309 0.534 0.640
02/23 21:57:06 - mmengine - INFO - Epoch(val) [80][625/625] coco/bbox_mAP: 0.4860 coco/bbox_mAP_50: 0.6560 coco/bbox_mAP_75: 0.5300 coco/bbox_mAP_s: 0.3090 coco/bbox_mAP_m: 0.5340 coco/bbox_mAP_l: 0.6400 data_time: 0.0003 time: 0.0170
请问有什么关于finetune复现方面的建议及指导吗?另外,期待finetune部分的完善及权重的更新~
The text was updated successfully, but these errors were encountered: