imageModelClass

[adeldaoud]

@cjerzak , I am checking the helper file of causalimages::AnalyzeImageConfounding and the argument imageModelClass but could not find instructions about thwat model classes are available to the user.

[cjerzak]

I'm thinking we should depreciate that option. We should probably just ViTs for everything to minimize code complexity. I will depreciated the CNN model class. We should likely just have (a) ViTs and (b) pre-trained models of whatever variant people would like to try.

[adeldaoud]

That is definitely one option. But I am thinking that it is hard to know what models users would like to apply to their analyses and for what reason. Implying that I wonder if we can keep model classes as an importable feature at least. E.g., there might be an application where uses would like to compare performance of various model classes on a problem -- applied or simulated -- and have several models would improve the versatility of causalimages.

Also, I recall you saying that we had something like seven different models, including pre-trained models,