/GluonBootcamp

Primary LanguageHTML

MXNet/Gluon Bootcamp

Setup

Since this repository contains submodules use the following command to clone:

git clone --recurse-submodules https://github.com/cyrusmvahid/GluonBootcamp.git

Schedule

Day 1

Title Description Duration Presenter
MXNet & Gluon Overview An overview of MXNet architecture and components 30 mins Thom
SageMaker Setup Get configured for the day ahead 15 mins Christian
Crash Course (LAB) Walkthrough of core Gluon components, and use them to create and train a convolutional neural network 120 mins Thom & Cyrus
Custom DataSets and DataLoaders Create customized DataSets and DataLoaders for various data types 30 mins Thom
MultiGPU Training (LAB) An introduction to training using multiple GPUs in Gluon, and hands on lab 60 mins Eden
Multi Host Training A deeper dive into cluster setup, sizing, optimization algorithms, and other considerations for distributed training 60 mins Eden & Cyrus
SageMaker Deep-Dive A deep-dive into SageMaker SDK and security model 30 mins Christian
SageMaker DevOps A deep dive into how to integrate SageMaker into development pipeline for teams of data scientists 30 mins Christian
Multi Host Training with SageMaker(LAB) Participants learn the simple steps to training using a cluster with Gluon and Amazon Sagemaker 30 mins Eden

Day2

Title Description Duration Presenter
Deep Learning AMI Setup An introduction to DLAMI and get configured for the day ahead 30 mins Christian
Beyond the Defaults Use simple regression model we take a look at alternatives for initialization, optimization and evaluation. 30 mins Cyrus
MXBoard Using MXNet with Tensorboard to monitor training 30 mins Thom
GluonNLP Walk through of key motivations and features of the GluonNLP toolkit. 30 mins Cyrus
Stacked Bidirectional LSTM (LAB) We work up from plain RNN to Stacked Bi-directional LSTMs using Gluon layers 60 mins Thom
LSTM Model (LAB) Participant in this section implement a simple LSTM network 60 mins Cyrus
LSTNet Theory This section describes theory of Long and Short Term Temporal Patterns with Deep Neural Networks 30 mins Cyrus
LSTNet Implementation (LAB) While referencing the paper, we construct the LSTNet model and train on multi-gpu 90 mins Thom
LSTNet Multi GPU (LAB) Scale training of LSTNet model with Multi GPU 30 mins Cyrus

Day3

Title Description Duration Presenter
Profiling MXNet Analysis of runtime code to identify performance bottlenecks 30 mins Thom
Moving LSTNet to SageMaker(LAB) The LSTNet code will be ported to train on Amazon SageMaker using the Python SDK 90 mins Eden
Distributed LSTNet on SageMaker(LAB) Participants will speed up the training of the LSTNet model with data distribution across multiple GPUs and hosts 30 mins Eden
SageMaker Automatic Model Tuning A look at SageMaker's hyperparameter optimization features 60 mins Thom
Deploying with SageMaker (LAB) We discover three scenarios, Deploying a model developed in SageMaker, Deploying model artefacts from S3, and deploying model artefacts from docker image 60 mins Eden
MXNet on Edge with Raspberry PI We demonstrate a simple model that runs on Raspberry PI 3 30 min Christian
MXNet on Edge with Amazon Greengrass IWe demonstrate deployment of a simple CV model on Amazon GreenGrass using Lambda 30 mins Christian
Keras-MXNet 2 (LAB) We use Keras to implement a bi-directional LSTM using Keras and the MXNet backend 30 mins Cyrus
GluonCV toolkit Walk through of key features of the GluonCV toolkit, with demo. 30 mins Thom

Day4

Title Description Duration Presenter
Gluon project structure Suggestions on how you can organize your MXNet Gluon projects. 30 mins Cyrus
Hackathon A project based on bi-directional LSTM will be implemented to incorporate all the material from the first three days. If participants have prepared datasets of significant size we attempt to optimize the distributed training for large batch sizes 6.5 hours All 
## Convert to markdown
# !notedown README.ipynb --to markdown > README.MD

## Convert to slides
# jupyter nbconvert <your notebook> --to slides --reveal-prefix reveal.js
## then copy the html file to `slides/note2slides` containing reveal.js