The training log of action classification models can not reproduce the paper's eval results

[Raion-Shin]

See swin-tiny's training log on K400, in videoswin's page, the top1 accuracy is about 73%, much lower than what is claimed in the paper's eval result, 78.9%. Does the training procedure really reproduce the paper?

2022/07/31 23:57:10 - mmengine - INFO - Epoch(val) [30][310/310]  acc/top1: 0.7306  acc/top5: 0.9053  acc/mean1: 0.7305
2022/07/31 23:57:10 - mmengine - INFO - The previous best checkpoint /mnt/lustre/daiwenxun/X/mmaction2/work_dirs/swin-t-2/best_acc/top1_epoch_28.pth is removed
2022/07/31 23:57:13 - mmengine - INFO - The best checkpoint with 0.7306 acc/top1 at 31 epoch is saved to best_acc/top1_epoch_31.pth.

@dreamerlin Thanks in advance for answering.