Is that normal that logits are not within -1 to 1?

openai / CLIP

CLIP (Contrastive Language-Image Pretraining), Predict the most relevant text snippet given an image

MIT License

24.55k stars 3.2k forks source link

Is that normal that logits are not within -1 to 1? #444

Closed Leo-T-Zang closed 2 months ago

Leo-T-Zang commented 3 months ago

Hi all,

I tried to train CLIP model on my dataset and reached over 0.8 top-1 accuracy. When I print all logits values out, I notice there are some values not within [-1,1] range, which is supposed to be the range of cosine similarity.

Is this normal or potentially I did anything wrong?

Thanks a lot!