Custom Dataset Valid Thresholding on Training With Skeleton Models

[Spawnfile]

Hello everyone,

I'm trying to train skeleton action models with my custom dataset with Pose3D-ntu60-2d and STGCN-ntu60-2d configs. On data extraction code which is provided on github repo of mmaction2 called ntu_pose_extraction.py I see a hardcoded resolution which is (1080, 1920) on frame shape and video res. My all videos are not in the same shape so I changed this as a dynamic assignment according the input videos shape which is sometimes (1080, 1920) and sometimes (1280, 720) etc. However after doing this the validation script (valid thresholding) that comes from mmengine ignore all videos of mine.

Additionally before changing dynamic resolution assignment, my about %70 videos of my dataset was ignored also.

here are my questions:

1. Should I hardcode this resolution section even if my videos are in different shapes ?
2. How should I prepare my dataset to get rid of the ignoration from mmengine, and what is the working idea of this valid thresholding ?

Cheers
Alper

[brian123214]

Were you able to come up with a solution? I also have the same problem.

[laurahgdrn]

Did you come up with a solution? I also get the same error.

[Spawnfile]

I'm not woking with Mmengine anymore, sorry to announce that.