Training hours & Imagenet accuracy

Hello, thanks for sharing your interesting work.

I was trying to reproduce the NesT-T ImageNet result in this [link](https://tensorboard.dev/experiment/AU4DxhjnRBieaPsgCWGxng/#scalars) using TPUs.

Here are my TPU-v3 8 cores result ([link](https://tensorboard.dev/experiment/6gzc6PWQRGamDWAubsOspA/#scalars&_smoothingWeight=0)) by using exactly the same hyperparameters in [imagenet_nest_tiny.py](https://github.com/google-research/nested-transformer/blob/main/configs/imagenet_nest_tiny.py)

As you can see, it takes 63 hours for training while your result takes 21 hours.
How can I reduce training hours such as your result?
If this difference came from the data loading time, could you tell me the types of data storage that you used?
Right now, I'm using the google cloud storage bucket for data storage.

Furthermore, I can see the accuracy difference around 0.5% (81.0 v.s. 81.5).
Could you explain this difference?