fail to reproduce result on paper

[yunfeixie233]

Hi,YiF: i'm trying to reproduce the result of RegProxy ViT-S/16 on ade20k. Following instruction in readme and use the correct pretrain checkpoint, I found the result is 46.22 on 80k iteration, which is far low than result on table 7 (47.2), could you please help me with this problem like providing the training log and seed?