Contextualized models combine deep learning and simpler graphical or linear models, inheriting the hyperparameters of both (and sometimes some extras). NOTMAD is a great example, with 6 separate groups of parameters.
We have usually approached this by grouping kwargs into dicts, but this often leads to overwriting the whole kwarg group when a single parameter in the dict needs to be re-specified. We need a new way of managing kwargs that still respects grouping, but also allows specifying individual parameters.
Contextualized models combine deep learning and simpler graphical or linear models, inheriting the hyperparameters of both (and sometimes some extras). NOTMAD is a great example, with 6 separate groups of parameters.
We have usually approached this by grouping kwargs into dicts, but this often leads to overwriting the whole kwarg group when a single parameter in the dict needs to be re-specified. We need a new way of managing kwargs that still respects grouping, but also allows specifying individual parameters.