-
Notifications
You must be signed in to change notification settings - Fork 5.5k
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
Add (Scheduled) Pseudo-Huber Loss training scripts to research projects #7527
Conversation
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
The kohya-ss scripts have been merged after throughout testing in kohya-ss/sd-scripts#1228. Will update the rest of this Pull Request soon and prepare for merging in Diffusers |
Sure. |
@sayakpaul updated all scripts, ready for merging |
Should I do anything about the code quality? |
you can run |
nvm |
Please ping me once the CI run is complete. |
@sayakpaul CI check ran, but failed. I'm getting exhaused 😢 |
If you check https://github.com/huggingface/diffusers/actions/runs/8641181505/job/23690277573?pr=7527, it clearly says the solution:
Is it not helpful? |
@sayakpaul Thanks for the instruction! Now I'll cross my fingers and hope it'll finally work🤞 Sorry for all these misunderstandings and extra commits 😓 (please, squash them when merging) |
Thank you for your contributions. |
…ts (#7527) * add scheduled pseudo-huber loss training scripts See #7488 * add reduction modes to huber loss * [DB Lora] *2 multiplier to huber loss cause of 1/2 a^2 conv. pairing of kohya-ss/sd-scripts@c6495de * [DB Lora] add option for smooth l1 (huber / delta) Pairing of kohya-ss/sd-scripts@dd22958 * [DB Lora] unify huber scheduling Pairing of kohya-ss/sd-scripts@19a834c * [DB Lora] add snr huber scheduler Pairing of kohya-ss/sd-scripts@47fb1a6 * fixup examples link * use snr schedule by default in DB * update all huber scripts with snr * code quality * huber: make style && make quality --------- Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
This PR adds an option to use (Scheduled) Pseudo-Huber loss from https://arxiv.org/abs/2403.16728 for text2image and Dreambooth trainings.
Because of this feature being experimental, it's firstly needed to add it to the research projects.
See #7488 for detailed description, plots and examples
@sayakpaul @kashif @patil-suraj