PyTorch Lightning — Let’s organize ML

dasayan05


1

Vote

Description:

The widespread success of machine (deep) learning in recent years has forced both academic researchers and industrial engineers to iterate over model/algorithm developments faster than ever before. Such fast-paced workflows have led to disorganised codebases that are hard to read, maintain and scale. PyTorch Lightning is a “batteries included” framework built on PyTorch that helps you write ML code that is organized, free of boilerplates, readable and scalable. By following a minimal “protocol”, one can leverage PyTorch Lightning to orchestrate and automate typical ML workflows that include training, evaluation, testing, inference, data/model versioning, logging, compute scaling etc. In this fully “example-based” talk, you will learn the basics of PyTorch Lightning as well as some of its advanced offerings — step by step.

A tentative flow of the talk would be the following:

  • Typical ML workflows & ‘bare-metal’ PyTorch
  • The core LightningModule & LightningDatamodule classes
  • A Trainer that orchestrates model & data
  • The “batteries” — saving/loading, versioning, logging, configuring

Prerequisites:

  • Python
  • General ML Knowledge
  • Basics of software development

Content URLs:

The contents of this talk will mostly be based on personal experiences with PyTorch Lightning for several years. The library references can be found in the official documentation of PyTorch Lightning.

Speaker Info:

Ayan holds a PhD in Deep Learning and currently works as a Senior AI Research Scientist at MediaTek Research UK. He is involved in both academic research and industrial engineering workflows where cutting-edge ideas are implemented, maintained and scaled in order to bring them to production. He has been using PyTorch Lightning nearly since its inception and recently started producing collaborative tutorials with the official PyTorch Lightning team. Please visit his personal website for information about his works, publications, talks and others.

Section: Artificial Intelligence and Machine Learning
Type: Talk
Target Audience: Beginner
Last Updated: