-
Notifications
You must be signed in to change notification settings - Fork 296
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
Cordiff usability and performance enhancements for custom dataset training #790
base: main
Are you sure you want to change the base?
Cordiff usability and performance enhancements for custom dataset training #790
Conversation
/blossom-ci |
/blossom-ci |
/blossom-ci |
/blossom-ci |
/blossom-ci |
Overall this is looking nice, thanks for doing a lot of cleanup and docstring/readme improvements! I checked for compatibility/overlap with other diffusion weather models (StormCast, ReGen) and it looks fine in that regard. I had a couple minor comments about the unit tests for patching utilities but other than that no major qualms. It would be good to address Akshay's comment about performance if we can, it does seem like avoiding allocating the full meshgrid is worth doing since the whole goal is to avoid too much memory allocation and ops at the full image resolution |
/blossom-ci |
@pzharrington thanks for your comments! I made the patching tests more robust and addressed @akshaysubr suggestions for performance improvements. |
/blossom-ci |
Modulus Pull Request
Description
This PR provides enhancements for usability of CorrDiff training on custom datasets. It includes:
Closes #726.
Closes #447.
Closes #417.
Closes #448.
Note
This PR introduces breaking changes on the way corrdiff config files are handled.
For training, older version of the config files is still supported.
For generation, a new parameter
generation.patching: true/false
is required.Checklist
Dependencies