Squeezing maximum performance from PyTorch models

In recent years, techniques such as 16-bit precision, accumulated gradients and distributed training have allowed models to train in record times.

In this, talk we will go through the implementation details of the 10 most useful of these techniques, including DataLoaders, 16-bit precision, accumulated gradients and 4 different ways of distributing model training across hundreds of GPUs. We will also show how to use these already built-in in PyTorch Lightning, a Keras-like framework for ML researchers.

William is the creator of PyTorch-Lightning and an AI PhD student at Facebook AI Research and NYU CILVR lab, advised by Kyunghyun Cho (who spoke at NYC Machine Learning two years ago). Before his PhD, he Co-founded AI startup NextGenVest (acquired by Commonbond). He also spent time at Goldman Sachs and Bonobos. He received his BA in Stats/CS/Math from Columbia University.

Note: You will need to bring your photo id to sign in downstairs. You also need to sign up with your full name, or contact one of the organizers if your full name isn't on your meetup profile.