Splitting and Sampling Strategies

[selmanozleyen]

The interface would be at least like the following: `AnnData -> dict[AnnData]'.

Let's limit our variables to consider to at most 2 (e.g., cell_type:a/b and split: train/val/test)

Cases:

• User wants to split for each set train, test, val s.t. all these preserve their cell type distribution. E.g. if there is quarter celltype a in the whole dataset in train, test and val they should also be 1/4 of their respective sets. (here no entry is duplicate but in sampling this might not be the case)
• User wants to sample each set so that the classes have equal proportion. (here there can be duplicate entries)

@FrancescaDr here do you have any more cases you'd like to discuss? Also which one would be more important for you?

[FrancescaDr]

Thanks for making the issue.

The first case describes exactly what I had in mind for creating train, val and test splits. I think this should happen in the init of the GraphAnnDataModule right?

Regarding the second case it sounds more something related to sampling during the data loader when you sample cells that should be predicted.

[selmanozleyen]

Ok, I think we can leave second strategy to later. Regarding GraphAnnDataModule, I think we would first to create a module called datamodules and add GraphAnnDataModule to the repository again.