BLIP2 Text Localization

[RozDavid]

Hey all,

First of all thanks for the cool project and the shared checkpoints. I was wondering if there is any way to extract attention maps with respect to all query tokens using the QFormer module. Theoretically it should still have a similar cross-attention module that BLIP had within the text encoder's base model, but I cant't find a way to access this information with normal callbacks.

All help is appreciated! Thanks a lot, David