Lightning tools

Customization on top of pytorch-lightning for quick and easy experimentations. Inspired by

  1. Speechbrain's mixed yaml configuration, dependency injection and code customization approach.
  2. Pytorch-lightning's Modular approach.
  3. AWS Cloudformation's Infrastucture-as-code

Principles

  • Plug-and-playable loss functions, models, data-modules and loggers.
  • Configuration-as-code
  • Pass by name, minimal pass by position
  • Functional and high composability
  • Dependency injection

TODO

  • Dataloader design is not as plug-and-play as expected
  • Grouped batching based on data length.
  • Think of a shorter name
  • Enable multi-gpu processing