search

NVIDIA / modulus

Open-source deep-learning framework for building, training, and fine-tuning deep learning models using state-of-the-art Physics-ML methods
https://developer.nvidia.com/modulus
Apache License 2.0
795 stars 172 forks source link

🚀[FEA]: CorrDiff training/inference without regression #478

Open jleinonen opened 2 months ago

jleinonen commented 2 months ago

Is this a new feature, an improvement, or a change to existing functionality?

New Feature

How would you describe the priority of this feature request

Medium

Please provide a clear description of problem you would like to solve.

For comparison purposes it would be good to have the ability to train CorrDiff with a plain diffusion model without regression.

Describe any alternatives you have considered

It's possible to do this by plugging in a "regression" model that returns all zeros but this is a rather hacky solution.

ram-cherukuri commented 2 months ago

@jleinonen Can you provide more context - If we dont need a regression step, wouldn't this be like training a diffusion model which Modulus already supports a AI researcher from doing?

jleinonen commented 2 months ago

@jleinonen Can you provide more context - If we dont need a regression step, wouldn't this be like training a diffusion model which Modulus already supports a AI researcher from doing?

@ram-cherukuri We have an example for a diffusion model but it doesn't have nearly all the functionality that the CorrDiff code has now. Whereas disabling the regression step in the CorrDiff codebase should be a matter of a few simple changes according to @mnabian.