Can GAFFO be used for the data hyper-cleaning example?

[dongboac]

Required prerequisites

• I have read the documentation https://boat.readthedocs.io/en/main/.
• I have searched the Issue Tracker to make sure this question has not already been answered.

Question

Hi, thanks for sharing this toolbox.

I was trying to use GAFFO with the data hyper-cleaning example, but I ran into an issue with the upper- and lower-level variables.

In this task, the upper-level variables seem to be the sample weights, while the lower-level variables are the model parameters. Since they have different shapes and structures, GAFFO does not seem to work directly.

Does the current implementation of GAFFO require the upper- and lower-level variables to match in shape? If so, does that mean it is not currently supported for the data hyper-cleaning example, or is there another way to set it up?

Thanks!

[QHofmann]

Thanks for your interest in BOAT.

You are right that data hyper-cleaning has different structures for upper-level variables (sample weights) and lower-level variables (model parameters). Currently, GAFFO in BOAT does not directly support this kind of variable transformation.

For this task, hyper-gradient based methods such as RAD/RGT may be more suitable. We will consider improving the support and documentation for such cases in the future.

Thanks for the suggestion!