Make a pruned version version
Is it technically feasible to make a smaller version of this model without losing to much accuracy?
I there anything I should take in account before going down this path?
does anyone who may read this, know of any metrics available on this?
This model is pretty old in ai terms but truly is amazing, I don't hear anyone talk about it. But I guess sdxl is going to change that soon...
https://github.com/LAION-AI/CLIP_benchmark metrics are there
Btw I will update CLIP benchmark soon with all models available in open clip, including the new datacomp and convnext ones.
About a smaller version, one option is distillation, I am not aware if someone did it already, but it is supported in open clip https://github.com/mlfoundations/open_clip#model-distillation.
@Kwissbeats in the meanwhile here are the full benchmark results if you want to have a look: https://gist.github.com/mehdidc/7968227624e13a0ca3875e4c05f966f1