CLIPScore

Hi,

Thank you for this interesting repository and contributions.
I clone the repo and tried to reproduce the results. However, I had higher values of CLIPScore than the paper claims. So, I undertake a little study to better understand the results of CLIPScore.
I used the clipa-ViT-bigG-14 pretrained on the DataComp-1B on different pairs of images :


Respectively : semantically different, same, ill-generated and mid- generated images (mid-generated means well generated but not perfectly)

In order to compute the ClipScore, I follow the provided pipeline (F.interpolate(), image_encode() and F.cosine_similarity).
For this pairs I have the following results : (approx. 10**-2)
- semantically different : 0,67
- same picture : 1.0
- ill-generated : 0,91
- mid-generated : 0,83
The remote sensing were generated with the pretrained weights for 8 inputs on the val_set with the configuration proposed (esrgan_baseline_8S2.yml), so the small esrgan proposed.

Thank you for your time,
Hope you can help me