Can not reproduce the split1 shot5 result

[tymanman]

I followed the guidance of README (without any change) to pretrain a base model and then randomly initilize the novel weights, finally finetune it.But I get a best result of nap of only 58.8，there is a big gap between my result and the result given in paper.How come

And I wanna konw whether the config files shown in the repo is the same as the ones you use in your experiment

[tymanman]

Recently, I repeat a lot of runs,finally get a better result of 60.16.

[kike-0304]

最近，我重复了很多次，终于得到了一个更好的结果，60.16。

Hi，how many GPUs did you use, thank you.