Fix logit scale calculation in zero_shot_iid.py

[work4cs]

bioclip/src/open_clip/model.py

Line 237 in 4ea3247

"logit_scale": self.logit_scale.exp()

[egrace479]

Back to how it had been ✅

[hlapp]

If this is the correct form, then we'll need to fix this in pybioclip as well.

And presumably this doesn't affect ordering after softmax()? But it does affect the reported probabilities?

[work4cs]

If this is the correct form, then we'll need to fix this in pybioclip as well.

And presumably this doesn't affect ordering after softmax()? But it does affect the reported probabilities?

I am thinking of if double exp(), some values might become too large and overflow. Then, the ranking becomes meaningless and wrong. And probabilities/confidence, yes, it might make the result seems over-confident.

[egrace479]

@work4cs is going to explore this more over the weekend before we make any changes.