Closed blengerich closed 2 years ago
Metamodels, datasets, and models seem to be fairly distinct types of regression objects. What would you think about making a regression directory, and splitting the types into their own files in the same regression directory?
That makes sense, especially because each model has attributes like "loss" and "dataloader" which require model-specific implementations but are universal to all model types.
Refactored in #60
The metamodels are model-specific so I think it makes sense to leave them in regression. Similarly the NOTMAD metamodel would go in networks. The correlation metamodel makes explicit use of the regression metamodels, so I also think it makes sense to import from regression there.
On this topic, it might be good to move regularization losses into regression since I think they're also built specifically for the regression metamodels.