Abstract: This tutorial introduces Gluon, a flexible interface that pairs MXNet’s speed with a user-friendly front end. Gluon can run as a fully imperative framework. In this mode, you enjoy native language features, painless debugging, and rapid prototyping. You can also deploy arbitrarily complex models with dynamic graphs. And when you need more performance, Gluon can also provide the speed of MXNet’s symbolic API by calling down to Gluon’s just-in-time compiler. In this crash course, we’ll cover the basics of how to code up deep learning models and deploy them on GPUs. We’ll cover the fundamentals of MXNet’s data structures, automatic differentiation, and walk through how to define neural networks at the atomic level, implementing each part yourself. Then we’ll introduce Gluon’s higher level abstractions for neural network layers, loss functions, and optimizers and show how you can implement CNNs, RNNs, and GANs with just a few lines of code. To go hands-on, bring a laptop with Python, MXNet and Jupyter installed and clone the tutorial repository http://github.com/zackchase/mxnet-the-straight-dope.
Bio: Zachary Lipton is a mad scientist at Amazon AI and assistant professor at Carnegie Mellon University (2018-). He researches ML methods, applications (especially to healthcare), and social impacts. In addition to corralling deep neural neurons and starting fires on Twitter (@zacharylipton), he is the editor of the Approximately Correct blog and lead author of Deep Learning - The Straight Dope, an interactive book teaching deep learning and MXNet Gluon through Jupyter notebooks.