MGDS is a custom dataset implementation for Pytorch that is built around the idea of a node based graph.
Each node in the graph can take inputs from previous nodes, do various calculations, and return outputs to nodes further down the chain. This modular approach enables highly complex data processing pipelines only by plugging together pre-built modules.
This project is currently developed for the use in OneTrainer, but it can be used in all kinds of different applications.