About the projection and prediction head dimension configs

[Dylan-H-Wang]

Hi,

Thank you for the amazing works and I am inspired a lot from it!

One thing I want to ask is how you designed the structure of your projection and predications heads, i.e., how to decide the number of layers and hidden and output dimensions? P.S. I notice that you have already mentioned the bottleneck structure of prediction MLP is helpful.

Thank you for the helps and time!