Support for weighted loss either per-batch or per-cell?

[kamilkrukowski]

Could scVI models support weighed loss and weighing individual cells or batch_key groups?

Some potential uses

• Downweighing over-represented atlases with low interesting variation.
• Improving ELBO on poorly represented subsets of interest i.e. rare cell types or small samples.

Comments To my knowledge, oversampling/undersampling is not a viable solution here because scvi-tools requires in-memory AnnData objects containing the entire dataset. Is there any way to "stream" data loading with meaningful oversampling that does not require holding all oversampled entries in memory for the entirety of training?

[canergen]

Hi, thanks for your question. @martinkim0 we wanted to have a stratified validation/train splitter at some time point but never followed up on this. I guess this would yield a very similar result. Currently you can subset your object before running scVI.

[canergen]

You can now used stratified train splitters in main (#2902, it will be released in scvi-tools 1.2). You can define a train set in which small cell-types are more abundant and will yield a similar increase of weights during training.