Training models with torch.Tensor input about scvi-tools HOT 4 OPEN

We cover the enhancement to use custom dataloader in the recent version of scVI-tools.
However, it is not clear yet which minimal checks (integer, gene names) we still want to perform.
About your example: @Intron7: Is this idea of having AnnData in torch recommended? What analysis capabilities are possible in this scenario? I thought this is meant to be done in rapids_singlecell. Does rapids copy back and forth between CPU and GPU or is the full data kept between processing steps on GPU?

from scvi-tools.

Thanks! Is there already a link to an example usage of this new version?
AFAIK rapids_singlecell keeps matrices on GPU without back and forth, (not using torch though) https://rapids-singlecell.readthedocs.io/en/latest/Usage_Principles.html

from scvi-tools.

We are still talking about how this would work. However at the moment whenever I use rsc I have to transform back to cpu and than use scvi. Rapids-singlecell really wants .X and .layers on the GPU so everything has to be in memory. I would really like if we used DLPack for this. DLPack allows for the 0 copy conversion from cupy and jax to torch.

from scvi-tools.

Hi @j-bac, thanks for the suggestion. We will be releasing a tutorial with our next release (v1.2) that covers a basic usecase with a custom dataloader. I'll note that we currently don't support inference methods yet (e.g. get_latent_representation), but it's something we're working on.

from scvi-tools.