[Bug] Unable to Replicate Reported AP for RTMPose

[HaziqRazali]

Prerequisite

• I have searched Issues and Discussions but cannot get the expected help.
• The bug has not been fixed in the latest version(https://github.com/open-mmlab/mmpose).

Environment

I’m trying to retrain from scratch to replicate the results of a specific model RTMPose-l, which was trained solely on the COCO WholeBody dataset. However, I’m unable to achieve the same performance as reported. According to the table, the expected whole AP is 61.1, but after training, I only get an AP of 54.59.

CUDA_VISIBLE_DEVICES=1 python tools/train.py projects/rtmpose/rtmpose/wholebody_2d_keypoint/rtmpose-l_8xb64-270e_coco-wholebody-256x192.py

I’ve attached a link to a google drive of the zip file containing the .pth and .log files. https://drive.google.com/file/d/1e9OPKBHD84mVNPr6YIzC59Z8ihD6l4Ye/view?usp=sharing

Am I missing something? Has anyone encountered a similar issue or have suggestions?

Reproduces the problem - code sample

CUDA_VISIBLE_DEVICES=1 python tools/train.py projects/rtmpose/rtmpose/wholebody_2d_keypoint/rtmpose-l_8xb64-270e_coco-wholebody-256x192.py

Reproduces the problem - command or script

CUDA_VISIBLE_DEVICES=1 python tools/train.py projects/rtmpose/rtmpose/wholebody_2d_keypoint/rtmpose-l_8xb64-270e_coco-wholebody-256x192.py

Reproduces the problem - error message

2025/04/06 19:17:25 - mmengine - INFO - Epoch(val) [270][3250/3254]    eta: 0:00:00  time: 0.069622  data_time: 0.006557  memory: 915  
2025/04/06 19:18:14 - mmengine - INFO - Evaluating CocoWholeBodyMetric...
2025/04/06 19:18:57 - mmengine - INFO - Epoch(val) [270][3254/3254]    coco-wholebody/AP: 0.545874  coco-wholebody/AP .5: 0.827304  coco-wholebody/AP .75: 0.603021  coco-wholebody/AP (M): 0.544161  coco-wholebody/AP (L): 0.578310  coco-wholebody/AR: 0.639767  coco-wholebody/AR .5: 0.884603  coco-wholebody/AR .75: 0.703243  coco-wholebody/AR (M): 0.618820  coco-wholebody/AR (L): 0.669342  data_time: 0.007104  time: 0.070287
2025/04/06 19:18:57 - mmengine - INFO - The previous best checkpoint /home/haziq/mmpose/work_dirs/rtmpose-l_8xb64-270e_coco-wholebody-256x192/best_coco-wholebody_AP_epoch_260.pth is removed
2025/04/06 19:18:58 - mmengine - INFO - The best checkpoint with 0.5459 coco-wholebody/AP at 270 epoch is saved to best_coco-wholebody_AP_epoch_270.pth.

Additional information

No response