[Feature] SCDL: Make PyTorch usage optional

[mlgill]

Problem & Motivation

SCDL seems to me to have two parts: a format for efficient access to row-wise data, and a PyTorch Dataset derivative.

The former has many uses beyond PyTorch and the project I work on explicitly tries to remain agnostic of a DL language. However, it seems that the primary entrypoint is SingleCellMemMapDataset and it uses a class that depends on torch Dataset.

It would be great if there were a lower level API that doesn't require PyTorch, thus making its usage and installation optional.

BioNeMo Framework Version

v2.6.3

Category

API/Interface

Proposed Solution

NA

Expected Benefits

Usability

Code Example