Positional encoding in multiple instance learning

[relyativist]

Hi Monai Universe,

I have took a closer look to mil tutorial and haven't found the positional encoding implementation of input sequence in case mil_mode is in ["att_trans", "att_trans_pyramid"]. Assuming WSI images are patched on dependent instances. I assume model would utilize the position of each instance to propagate spatial information. From this assumption, I have to questions:

1. Do I get it right that in current tutorial setup, without positional encoding, we assume that multi-head attention doesn't incorporate positional information, just each instance tile embedding representation.
2. From theoretical side, does adding the positional encoding make sense in case of WS Images?

Thanks in advance!

[relyativist]

Can anyone anyone elaborate on this?