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
{{ message }}
This repository has been archived by the owner on Sep 19, 2020. It is now read-only.
You could increase GPT 3 accuracy by using Ranger, which combine state of the art optimizers + gradient centralization https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer
You seem to be using the Adam optimizer. It has been succeeded by RAdam (rectified Adam). Ranger will bring you this improvment and a lot more synergistic others, for free.
Hortogonally, you would probably benefit from Mish too instead of the one you use (Relu ?) but should be tested after Ranger as it could regress accuracy (even if unlikely) https://github.com/digantamisra98/Mish
The text was updated successfully, but these errors were encountered:
You could increase GPT 3 accuracy by using Ranger, which combine state of the art optimizers + gradient centralization
https://github.com/lessw2020/Ranger-Deep-Learning-Optimizer
You seem to be using the Adam optimizer. It has been succeeded by RAdam (rectified Adam). Ranger will bring you this improvment and a lot more synergistic others, for free.
Hortogonally, you would probably benefit from Mish too instead of the one you use (Relu ?) but should be tested after Ranger as it could regress accuracy (even if unlikely)
https://github.com/digantamisra98/Mish
The text was updated successfully, but these errors were encountered: