Sergei Rybakov
Sergei Rybakov
Also there is already a loader with built-in converter to pytorch tensors https://github.com/theislab/anndata/tree/master/anndata/experimental/pytorch https://anndata.readthedocs.io/en/latest/api.html#experimental-api
I agree, a different thing if you have this in mind. But then it gets unclear if things like pytorch (or jax) tensors should be treated in loaders or directly...
Hm, i think the zero copy conversion can be done via `dlpack`: https://jax.readthedocs.io/en/latest/jax.dlpack.html https://pytorch.org/docs/stable/dlpack.html
Yes, the way pytorch lightning tries to inject the sampler is not supported, probably you need to disable this somehow. But i will check what can be done to fix...
@mbuttner thank you, i will check their proposed changes.
Hi, missed this, i will check.
As for `num_workers`, yes, it is not implemented and also makes no sense to use parallel gathering of indices and then collating for an `AnnData` or `AnnCollection` object, the batch...
Yes, thanks, you are right.
Agree, partial access would be really useful. Also the current backed mode isn't just incomplete, it is partially broken wrt subsets etc.
We don't require flax, it is probably scvi-tools installation error. Are you able to install scvi-tools separately?