ImageNet evaluation with linear-probe - logistic regression's C parameter

[IdoGalil]

Thank you for your amazing paper,

I am trying to evaluate CLIP with a linear-probe on ImageNet, but wish to save some of the compute needed for the sweep required to optimize the C hyperparameter for each CLIP model. Is this possible for you to provide the values used for your models on ImageNet?

You also mentioned on another thread you might be able to share the classification head you used: https://github.com/openai/CLIP/issues/12 This could also be a great help, but even the values for the C hyperparameter would be very useful.

Thanks