Cannot reproduce the pretraining experiment on K400 dataset

[Raion-Shin]

Hi, I tried to reproduce the swin-tiny pretraining with the config you provided, but the validation result doesn't match your paper's result at all. The result I get is 55.28% top1-acc and 79.28% top5-acc. Is there anything missed?

[innat]

I translated to keras 3 and using torch backend, the score seems well.

 Acc1=77.690, Acc5=93.297

Notebook.

You probably should check your dataloader and frame size and normalizaiton stuff.