google

/

owlvit-base-patch32

like 107

Hello,
I am wondering how can I ensemble multiple text/image queries?

I assume there are two possibilities.

1. Averaging probs(= sigmoid of logit) of different queries
2. Plug-in averaged token to model

From my perspective,
Text-ensemble should be done with 1’s manner and image-ensemble for 2.
(bc I think averaging text token could mess thr embedding)

Please share your opinion regarding this!