mli / cvpr17

CVPR 2017 Tutorial

An Introduction to MXNet/Gluon

Note: The new version is availabe at:

• English: http://gluon.mxnet.io/
• 中文: https://zh.gluon.ai/

MXNet is widely used in production environments owing to its strong reputation for speed. Now with gluon, MXNet’s new imperative interface, doing research in MXNet is easy.

In this tutorial, we will walk through how to use gluon to implement various algorithms. We will present every concept in details, no deep learning background is required to attend. We encourage the audience to bring their laptops to have a hands-on experience with gluon.

This tutorial is on 9AM--12AM, 7/26 Wed, at 315, convention center. The detailed schedule is as follows:

1. 9:00 - 9:30: What is Gluon, why? slides
2. 9:30 - 10:30: Part I:
   • 15min: n-dimensional arrays notebook
   • 15min: auto differentiation notebook
   • 15min: linear regression notebook I, notebook II
   • 15min: convolution neural networks notebook I, notebook II
3. 10:30 - 11:00: Coffee break
4. 11:00 - 12:00: Part II:
   • 10min: hybridizing imperative and symbolic programming notebook
   • 10min: MXNet backend engine slides
   • 15min: multi-GPU and multi-machine slides
   • 15min: training with multi-GPUs notebook I, notebook II
   • 5min: training with multi-machines notebook

Note: all notebooks are runnable, the setup instructions are available here

cvpr17's Issues

How can I do physically mini-batch training with gluon?

Hi Mu, I read the tutorials, examples, and source code of gluon. It's very interesting and easy to use. But my question is how to do mini-batch with dynamic computation graphs such as TreeLSTM. I don't think the MXNet example tree_lstm really does mini-batch physically. The example calculates forward() and backward() for every instance, then it just update gradients to weight after a batch_size, but in the core engine, the real calculation with CPU/GPU is not parallelized by mini-batch. Do I understand it right?
So does mxnet have a plan on how to batch dynamic computation graphs? Like TensorFlow Fold.