[MaxViT] Block/Grid Attention question

[sonderlau]

Hi, thank you for your great work. I've seen all lines of your MetNet 3's implementation. Reviewing from your codes is joyful, and i have learned so much.
However, when I was checking the MaxViT module, and read the whole original paper, I had a question about the Block/Grid Attention design.

In appendix, formula (4) and (5), the final shape of the block and grid attention are the same: (HW/(P*P), P*P, C) (since P = G), only the grid did a swapaxes, but i dont think that's a big deal.

And in your MaxViT's implementation
from line 259 to line 267, these 2 attentions implementations are the same codes.
Also, I've checked offical implementation, nothing special.

So, from the paper's shape description and implementations, there's no sign of local attention and global attention, I'd love to call these as 2 local attentions stack.

I'm a new learner to DL, above all is my humble thoughts. Please leave comments if you have any ides.Thank you for your time.