Questions on CLIP calibration

[DianCh]

Hi, I was wondering how you calculated the confidence for CLIP models with their zero-shot predictions. Specifically:

• Did you take the softmax of the logits calculated by the following snippet (taken from the original CLIP paper)?
• If not, could you help clarify what values are used as confidence?

Thank you for the wonderful work of the paper!

[josipd]

Hello,

our instantiation of the CLIP model is checked in here: https://github.com/google-research/robustness_metrics/blob/5a380c4ec11fa85255b5db643c64efa42d749110/robustness_metrics/models/clip.py