Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Feat]: Cosine with hard restarts Tunable floor #546

Open
ppbrown opened this issue Nov 6, 2024 · 3 comments
Open

[Feat]: Cosine with hard restarts Tunable floor #546

ppbrown opened this issue Nov 6, 2024 · 3 comments
Labels
enhancement New feature or request

Comments

@ppbrown
Copy link
Contributor

ppbrown commented Nov 6, 2024

Describe your use-case.

How about adding a "..." menu option for CosHard, that allows the user to set a floor, instead of going to 0?
I'm not sure what type of value would be best.. maybe just a specific LR value, where if it hits THAT value, reset to full, instead of dropping all the way to virtual-0 before the reset?

What would you like to see as a solution?

LR floor value option

Have you considered alternatives? List them here.

?

@ppbrown ppbrown added the enhancement New feature or request label Nov 6, 2024
@ppbrown ppbrown changed the title [Feat]: Cosine with hard restarts Tunable [Feat]: Cosine with hard restarts Tunable floor Nov 6, 2024
@Zokreb
Copy link

Zokreb commented Nov 6, 2024

Hi community,
I would second this idea :)

@Calamdor
Copy link
Contributor

Calamdor commented Nov 6, 2024

It is possible to do with the custom learning rate scheduler.
https://github.com/Nerogar/OneTrainer/wiki/Custom-Scheduler#cosine-annealing

@ppbrown
Copy link
Contributor Author

ppbrown commented Nov 6, 2024

Thanks for the tip. But... if I wanted to mess around with things like manually filling in long.code.magic.function.names like that, I wouldnt be using OneTrainer :-/

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
enhancement New feature or request
Projects
None yet
Development

No branches or pull requests

3 participants