The reproducibility of the results on ImageNet-1K in the paper

[stareone]

I carefully implement this code but fail to reproduce the results reported in the paper (https://openreview.net/forum?id=JAezPMehaUu). As claimed in the paper, MosRep achieves 76.2% linear probing accuracy and 70.4% kNN accuracy. However, I implement MosRep using this code and only obtain 72.3% linear probing accuracy and 61.3% kNN accuracy on ImageNet-1K. There is a significantly large gap between the reported result and the reproduced one. I also comfirm that the settings of this paper are roughly identical to the ones in the paper. How can I reproduce the results reported in the paper?

Thanks!