Possible bug in utils_training/evaluation.py

[JunbongJang]

I implemented the training code myself, but my trained model yielded PCK less than 10 at the threshold 0.1. Then, I found that the function, eval_kps_transfer_test() is using source keypoints as ground truth instead of target keypoints. After the fix, I get a PCK of about 42 at the threshold 0.1

Inside the function eval_kps_transfer_test(cls, prd_kps, batch, alpha), batch['src_kps'] should be changed to batch['trg_kps']

Could you verify whether this is right?

Thank you.

[SunghwanHong]

I believe that may be the case.

Many times when the results were strangely bad, I first try to swap the trg and src.

[JunbongJang]

Ok, thank you. Could you provide training hyperparameters, such as batch size and training epochs?

[SunghwanHong]

I can't really remember the exact setting except for what is stated in the paper. Apologies for the inconvenience.

By the way, I updated the pretrained weights link.

[JunbongJang]

Ok, training hyperparameters are inside the "args.pkl" file in your updated link. Thank you.