You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In the recent ML news, I noticed mentions of FasterViT, a new model with some architecture changes that is said to outperform others in the performance to accuracy ratio. I immediately wondered if it can give tools like this any performance gain.
Integrating this model will require training a CLIP with it from scratch. This sounds like a lot of fun and should be straightforward-ish to implement if we take the open_clip codebase, but it is out of the closest roadmap for this repo. Closing now with the possibility of circling back to this in the future.
In the recent ML news, I noticed mentions of FasterViT, a new model with some architecture changes that is said to outperform others in the performance to accuracy ratio. I immediately wondered if it can give tools like this any performance gain.
https://github.com/NVlabs/FasterViT
The text was updated successfully, but these errors were encountered: