Skip to content

Principles behind nn-template

When developing neural models ourselves, we often struggled with:

  • Reproducibility. We strongly believe in the reproducibility requirement of scientific work.
  • Framework Learning. Even when you find (or code yourself) the best framework to fit your needs, you still end up in messy situations when collaborating since others have to learn to use it;
  • Avoiding boilerplate. We were bored to write the same code over and over in every project to handle the typical ML pipeline.

Over the course of the years, we fine-tuned our toolbox to reach this local minimum with respect to our self-imposed requirements. After many epochs of training, the result is nn-template.

nn-template is not a framework

  • It does not aim to sidestep the need to write code.
  • It does not constrain your workflow more than PyTorch Lightning does.