Torch NOTMAD is a big improvement, but to phase out the tensorflow version we need better integration and consistent style with the regression module. Most notably, NOTMAD is reliant on a tensor device switch during prediction to be device agnostic and the NOTMAD datamodule doesn't match the regression dataloader. We should figure out which type of dataset is preferred for Contextualized. Remaining issues are mostly lack of documentation/clarity and code duplication.
Resolved by #129 #131 and #132. Still need to review data handling patterns, but this is an issue relates to regression as well and goes beyond NOTMAD.
Torch NOTMAD is a big improvement, but to phase out the tensorflow version we need better integration and consistent style with the
regression
module. Most notably, NOTMAD is reliant on a tensor device switch during prediction to be device agnostic and the NOTMAD datamodule doesn't match the regression dataloader. We should figure out which type of dataset is preferred for Contextualized. Remaining issues are mostly lack of documentation/clarity and code duplication.