Patch-based CorrDiff Integration - 2

[tge25]

Modulus Pull Request

Description

Add the option to incorporate the regression's output as an additional conditioning of the diffusion model. Allow arbitrary length of sinusoid positional embedding. Enable learnable positional embedding. Move positional embedding to network architecture Update the config file for patch-based training and generation.

Checklist

• [x] I am familiar with the Contributing Guidelines.
• [x] New or existing tests cover these changes.
• [ ] The documentation is up to date with these changes.
• [x] The CHANGELOG.md is up to date with these changes.
• [ ] An issue is linked to this pull request.

Dependencies

[mnabian]

/blossom-ci

[mnabian]

@tge25 CI failed, please black format your code

[tge25]

Thanks for adding this important feature into main @tge25.

I'm finding the flow of the configuration a bit hard to follow since there are a lot of new conditional statements added. It also seems like there is too much coupling in the code since you needed to add new if statements in the loss, training_loop and datasets.

A feature like positional encoding feature should only need to touch one place in the code ideally. I think this will be less tightly coupled, if we move more of the position encoding logic into the architecture.

Also, doesn't seem like there any new unit tests of the changes to modulus/

Thanks @nbren12 for all the comments. I have cleaned up the code and moved the positional embedding to network architecture. Some unit tests have also been added.

[mnabian]

/blossom-ci

[tge25]

/blossom-ci

[mnabian]

/blossom-ci

[tge25]

/blossom-ci

[mnabian]

/blossom-ci

[tge25]

/blossom-ci

[tge25]

/blossom-ci

[tge25]

/blossom-ci

[mnabian]

/blossom-ci