MCL303-Deep Learning with Apache MXNet and Gluon

© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
AWS re:INVENT
Sunil Mallya
Miro Enev
S r . A I S o l u t i o n s A r c h i t e c t , A W S D e e p L e a r n i n g
M C L 3 0 3 - D e e p L e a r n i n g w i t h A p a c h e M X N e t a n d G l u o n
S r . S o l u t i o n A r c h i t e c t , N V I D I A D e e p L e a r n i n g
M C L 3 0 3
N o v e m b e r 2 7 , 2 0 1 7

Agenda
• Introduction to AI at Amazon
• AI at scale: challenges and solutions
• Model deployment
• Visualizing deep learning
• Apache MXNet overview
• What is Gluon?
• Live coding a neural network with Gluon
• Customer stories with Apache MXNet
• Getting started and Apache MXNet resources

Amazon ML Stack

One-click deep learning
Up to~40k CUDA cores
Apache MXNet
TensorFlow
Keras 1.2, 2.0
Caffe, Caffe2
Torch, Pytorch
CNTK
Pre-configured CUDA drivers
Anaconda, Python2.7, Python3
+ AWS CloudFormation template
+ Container image
Deep Learning AMI
https://bit.ly/deepami

The “Learning” in Deep Learning
0.4 0.3
0.2 0.9
...
backpropagation (gradient descent)
Y1 != Y
0.4 ± 𝛿 0.3 ± 𝛿
new
weights
new
weights
0
1
0
1
1
.
.
--
Y
input
label
...
Y1

Deep Learning in Computer Vision
.02
.85
p(cat)
p(dog)
Convolutional neural
network
Layer 1 Layer 2 Output
Explore spatial information with convolution layers
Img src: https://leonardoaraujosantos.gitbooks.io/artificial-inteligence

Deep Learning in Natural Language
Processing
Recurrent
neural networks
input
output
state
Variable length input and output sequences

AI and Deep Learning Story

DL IN CONTEXT
AI
ML
DL
AI

ML
TRIBES

EVOLUTIONIS
T
CONNECTIONI
ST
BAYESIAN
ANALOGISTSYMBOLIST
BOOSTER
Generalization Performance

Learning at Scale

Native distributed training supported
Supports distributed training on multiple CPU/GPU machines to take advantage of cloud scal
Flexible programming model
Supports both imperative and symbolic programming maximizing efficiency and productivity
Portable from the cloud to the client
Runs on CPUs or GPUs, on clusters, servers, desktops, or mobile phones
Multi-lingual | No need to learn a new language
Python, R, Scala, Julia, C++, Matlab, or Javascript
Performance 0ptimized
Optimized C++ backend engine parallelizes both I/O regardless of source language

TESLA V100
THE MOST ADVANCED DATA CENTER GPU EVER BUILT
5,120 CUDA cores [ 640 Tensor cores ]
7.5 FP64 TFLOPS | 15 FP32 TFLOPS
120 Tensor TFLOPS
16GB HBM2 @ 900 GB/s
300 GB/s NVLink

Distributed Training

Ingest Next Batch, with Updated Model

Demo

Apache MXNet
Single implementation of
backend system and
common operators
Performance guarantee
regardless of which
front-end language is
used
Front end
Backend

Deep Learning Framework Comparison
Apache MXNet TensorFlow Cognitive Toolkit
Industry Owner
N/A – Apache
Community
Google Microsoft
Programmability
Imperative and
Declarative
Declarative only Declarative only
Language
Support
R, Python, Scala, Julia,
Cpp. Javascript, Go,
Matlab and more..
Python, Cpp.
Experimental Go and
Java
Python, Cpp,
Brainscript.
Code Length |
AlexNet (Python)
44 sloc 107 sloc using TF.Slim 214 sloc
Memory Footprint
(LSTM)
2.6GB 7.2GB N/A
*sloc – source lines of code

Introducing Gluon
Simple, easy-to-
understand code
Flexible, imperative
structure
Dynamic graphs
High performance
 Neural networks can be defined using simple, clear, concise code
 Plug-and-play neural network building blocks—including predefined layers,
optimizers, and initializers
 Eliminates rigidity of neural network model definition and brings together the
model with the training algorithm
 Intuitive, easy-to-debug, familiar code
 Neural networks can change in shape or size during the training process to
address advanced use cases where the size of data feed is variable
 Important area of innovation in natural language processing (NLP)
 There is no sacrifice with respect to training speed
 When it is time to move from prototyping to production, easily cache neural
networks for high performance and a reduced memory footprint

Let’s Build Our First Gluon
Application

# code from https://github.com/zackchase/mxnet-the-straight-dope
num_fc = 512
net = gluon.nn.Sequential()
with net.name_scope():
net.add(gluon.nn.Conv2D(channels=20, kernel_size=5, activation='relu'))
net.add(gluon.nn.MaxPool2D(pool_size=2, strides=2))
net.add(gluon.nn.Conv2D(channels=50, kernel_size=5, activation='relu'))
net.add(gluon.nn.MaxPool2D(pool_size=2, strides=2))
# The Flatten layer collapses all axis, except the first one, into one axis.
net.add(gluon.nn.Flatten())
net.add(gluon.nn.Dense(num_fc, activation="relu"))
net.add(gluon.nn.Dense(num_outputs))
# Handy class available here
https://github.com/sunilmallya/dl-twitch-series/blob/master/cnn_mnist_gluon_simplified.ipynb
CNN in Gluon

data = mx.symbol.Variable('data')
# first conv layer
conv1 = mx.sym.Convolution(data=data, kernel=(5,5), num_filter=20)
tanh1 = mx.sym.Activation(data=conv1, act_type=”relu")
pool1 = mx.sym.Pooling(data=tanh1, pool_type="max", kernel=(2,2), stride=(2,2))
# second conv layer
conv2 = mx.sym.Convolution(data=pool1, kernel=(5,5), num_filter=50)
tanh2 = mx.sym.Activation(data=conv2, act_type=”relu")
pool2 = mx.sym.Pooling(data=tanh2, pool_type="max", kernel=(2,2), stride=(2,2))
# first full connected layer
flatten = mx.sym.Flatten(data=pool2)
fc1 = mx.symbol.FullyConnected(data=flatten, num_hidden=500)
tanh3 = mx.sym.Activation(data=fc1, act_type=”relu")
# second full connected layer
fc2 = mx.sym.FullyConnected(data=tanh3, num_hidden=10)
# softmax loss
lenet = mx.sym.SoftmaxOutput(data=fc2, name='softmax')
Convolutional Neural Net i n Symbolic MXNet

Object Detection Using SSD
https://github.com/zhreshold/mxnet-ssd https://github.com/zackchase/mxnet-the-straight-dope/
S
Y
M
B
O
L
I
C
G
L
U
O
N

Train with default settings:
python -m sockeye.train
--source train-corpus.de
--target train-corpus.en
--validation-source dev-corpus.de
--validation-target dev-corpus.en
--output model-dir
Decode with default settings:
python -m sockeye.translate
--models model-dir
# Sockeye Framework Code
https://github.com/awslabs/sockeye
Sequence to Sequence
<BOS> la casa blanca
the white house <EOS>

https://github.com/apache/incubator-mxnet/tree/master/example/recommenders
Recommender Systems
source: Jacob Schreiber

Generative Adversarial Networks (GAN)
[
.76
.14
.83
-.06
]
Z
“Fake” image
“Real” image
“Fake” image
https://github.com/apache/incubator-mxnet/tree/master/example/

Customer Stories with Apache MXNet

SigOpt uses deep learning in its optimization
platform to configure and improve machine
learning and AI models for its clients
• Improves accuracy and lowers training costs
using Apache MXNet and other frameworks
running on Amazon EC2 P2 and SigOpt HPO
• Dramatically reduces computation costs by
eight times or more compared to traditional
optimization methods
Optimizing AI Models for Better, Faster Results

Autonomous Driving

Using Deep Learning to Drive Targeted Advertising
GumGum uses deep learning to provide actionable
insights to help advertisers create highly visible and
targeted campaigns
• Mantii, GumGum’s real-time social media listening
tool built on deep learning, can analyze the more
than 1.8 billion social images posted daily, where
80% of them lack text to help marketers find them
• With deep learning on AWS using Apache MXNet
and Amazon EC2 P2 instances, the company
reduced model training time from 21 hours to 6
hours

Apache MXNet Cheat Sheet
Cheat sheet: bit.ly/2xTIwuj

Thank you!
s m a l l y a @ a m a z o n . c o m
m e n e v @ a m a z o n . c o m

MCL303-Deep Learning with Apache MXNet and Gluon

Empfohlen

Empfohlen

Weitere ähnliche Inhalte

Was ist angesagt?

Was ist angesagt? (20)

Ähnlich wie MCL303-Deep Learning with Apache MXNet and Gluon

Ähnlich wie MCL303-Deep Learning with Apache MXNet and Gluon (20)

Mehr von Amazon Web Services

Mehr von Amazon Web Services (20)

MCL303-Deep Learning with Apache MXNet and Gluon