Implement metrics for evaluating image quality

[soumik12345]

• [ ] Fréchet inception distance
• [ ] Inception Score
• [ ] Kernel Inception Distance
• [ ] Universal Image Quality Index
• [ ] CLIP Maximum Mean Discrepancy

[soumik12345]

Hi @sayakpaul, can you please check the way I'm computing FID in this PR before I proceed to add the other metrics?

Here's a sample evaluation trace on a small subset 10 datapoints from the coco validation set https://wandb.ai/geekyrakshit/t2i_eval/weave/calls/1d34168d-bf4e-4364-908b-f1b323152ce6?tracetree=1

[soumik12345]

Hi @sayakpaul, can you please check the way I'm computing FID in this PR before I proceed to add the other metrics?

Here's a sample evaluation trace on a small subset 10 datapoints from the coco validation set https://wandb.ai/geekyrakshit/t2i_eval/weave/calls/1d34168d-bf4e-4364-908b-f1b323152ce6?tracetree=1

I am closing this PR, because I made one mistake: metrics like FID are calculated between distributions and not between individual images.

[sayakpaul]

@soumik12345 LMK when you have fixed the issues you found and have opened a new PR.

[sayakpaul]

Also, we don't need to report all the metrics here I think they are quite outdated TBH. No image generation paper from the recent days have reported them as per my knowledge and observation. So, FID and CMMD are fine.

Also, don't land all the metrics in a single PR. Let's tackle them one by one, please.