I’m trying to train using data_modules passed to a model’s .train, because I already have a train / valid split that I need to respect, and don’t want SCVI to redo the split internally.
For some reason, there isn’t a built-in way to do this as far as I can tell, to just specific two anndatas or to specify indices over a single anndata for training and validation?
But now, I need the anndata manager as SCVI.setup_anndata does a lot of work to setup the batch that the model can ingest, and the anndata manager is needed to create the dataset object that is then passed to dataloaders, following the batch format set by setup_anndata, from my understanding. I’m taking inspiration from scvi.dataloaders.DataSplitter, which is what scvi-tools uses internally when one just calls model.train scvi-tools/scvi/dataloaders/_data_splitting.py at 95f2e1d2fa921c6433a04e257d844280ba0c25e5 · scverse/scvi-tools · GitHub.
Currently, I don’t see a clean way to get the anndata manager. It would be so much more simple if SCVI.setup_anndata(anndata) returned the anndata manager directly.