Large Scale Distributed Systems for

Training Neural Networks

Jeff Dean & Oriol Vinyals
Google
Google Brain team in collaboration with many other teams
 Background
Google Brain project started in 2011, with a focus on
pushing state-of-the-art in neural networks. Initial
emphasis:

● use large datasets, and

● large amounts of computation

to push boundaries of what is possible in perception and

language understanding
 Overview
● Cover our experience from past ~5 years
○ Research: speech, images, video, robotics, language understanding,
NLP, translation, optimization algorithms, unsupervised learning, …

○ Production: deployed systems for advertising, search, GMail, Photos,

Maps, YouTube, speech recognition, image analysis, user prediction, …

● Focus on neural nets, but many techniques more

broadly applicable
 Overview
● Demonstrate TensorFlow, an open source machine
learning system
○ Our primary research and production system
○ Show real examples
○ Explain what’s happening underneath the covers
 Outline

● Introduction to Deep Learning

● TensorFlow Basics
○ Demo
○ Implementation Overview
● Scaling Up
○ Model Parallelism
○ Data Parallelism
○ Expressing these in TensorFlow
● More complex examples
○ CNNs / Deep LSTMs
 Growing Use of Deep Learning at Google
# of directories containing model description files Across many
products/areas:
Android
Unique Project Directories

Apps
drug discovery
Gmail
Image understanding
Maps
Natural language
understanding
Photos
Robotics research
Speech
Translation
YouTube
… many others ...

Time
Deep Learning
Universal Machine Learning
Speech Speech
Text Text
Search Search
Queries Queries
Images Images
Videos Videos
Labels Labels
Entities Entities
Words Words
Audio Audio
Features Features
Deep Learning
Universal Machine Learning
...that works better than the alternatives!
Current State-of-the-art in:
Speech Recognition
Image Recognition
Machine Translation
Molecular Activity Prediction
Road Hazard Detection
Optical Character Recognition
...
ConvNets
Some More Benefits

Deals very naturally w/sequence data (text, speech, video...)

Very effective at transfer learning across tasks

Very easy to get started with a commodity GPU

A common ‘language’ across great many fields of research

Two Generations of Distributed ML Systems

1st generation - DistBelief (Dean et al., NIPS 2012)

● Scalable, good for production, but not very flexible for research

2nd generation - TensorFlow (see tenorflow.org and

whitepaper 2015, tensorflow.org/whitepaper2015.pdf)
● Scalable, good for production, but also flexible for variety of research uses
● Portable across range of platforms
● Open source w/ Apache 2.0 license
Need Both Large Datasets & Large, Powerful Models
“Scaling Recurrent Neural Network Language Models”, Williams et al. 2015
arxiv.org/pdf/1502.00512v1.pdf
 Large Datasets + Powerful Models
● Combination works incredibly well
● Poses interesting systems problems, though:
○ Need lots of computation
○ Want to train and do experiments quickly
○ Large-scale parallelism using distributed systems
really only way to do this at very large scale
○ Also want to easily express machine learning ideas
 Basics of Deep Learning
● Unsupervised cat
● Speech
● Vision
● General trend is towards more complex models:
○ Embeddings of various kinds
○ Generative models
○ Layered LSTMs
○ Attention
Learning from Unlabeled Images

• Train on 10 million images (YouTube)

• 1000 machines (16,000 cores) for 1 week.
• 1.15 billion parameters
 Learning from Unlabeled Images

Optimal stimulus
by numerical optimization
Top 48 stimuli from the test set
 Learning from Unlabeled Images

Optimal stimulus
by numerical optimization
Top 48 stimuli from the test set
 Adding Supervision

Top stimuli for selected neurons.

 Speech: Feedforward Acoustic Models
Model speech frame-by-frame,
independently

Simple fully-connected networks

Deep Neural Networks for

Acoustic Modeling in Speech
Recognition
Hinton et al. IEEE Signal
Processing Magazine, 2012
CLDNNs
Model frequency invariance using 1D convolutions

Model time dynamics using an LSTM

Use fully connected layers on top to add depth

Convolutional, Long Short-Term Memory,

Fully Connected Deep Neural Networks
Sainath et al. ICASSP’15
 Trend: LSTMs end-to-end!
Speech Acoustics Phonetics Language Text

Train recurrent models that also incorporate Lexical and Language Modeling:

Fast and Accurate Recurrent Neural Network

Acoustic Models for Speech Recognition, H. Sak et al. 2015

Deep Speech: Scaling up end-to-end speech recognition, A. Hannun et al. 2014

Listen, Attend and Spell, W. Chan et al. 2015

 CNNs for Vision: AlexNet

ImageNet Classification with Deep Convolutional Neural Networks

Krizhevsky, Sutskever and Hinton, NIPS 2012
The Inception Architecture (GoogLeNet, 2015)
Basic module, which is then
replicated many times
The Inception Architecture (GoogLeNet, 2015)

Going Deeper with Convolutions

Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov,
Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich

ArXiv 2014, CVPR 2015

Inception-v3 (December 2015)

http://arxiv.org/abs/1512.00567
 Rapid Progress in Image Recognition
Team Year Place Error (top-5) Params ImageNet
XRCE (pre-neural-net explosion) 2011 1st 25.8% challenge
classification
Supervision (AlexNet) 2012 1st 16.4% 60M
task
Clarifai 2013 1st 11.7% 65M

MSRA 2014 3rd 7.35%

VGG 2014 2nd 7.32% 180M

GoogLeNet (Inception) 2014 1st 6.66% 5M

Andrej Karpathy (human) 2014 N/A 5.1% 100 trillion?

BN-Inception (Arxiv) 2015 N/A 4.9% 13M

Inception-v3 (Arxiv) 2015 N/A 3.46% 25M

Models with small number of parameters fit easily in a mobile app (8-bit fixed point)
Today’s News: Pre-trained Inception-v3 model released
http://googleresearch.blogspot.com/2015/12/how-to-classify-images-with-tensorflow.html

Dear TensorFlow community,

Today we are releasing our best image classifier trained on ImageNet data. As described in our
recent Arxiv preprint at http://arxiv.org/abs/1512.00567, an ensemble of four of these models
achieves 3.5% top-5 error on the validation set of the ImageNet whole image ILSVRC2012
classification task (compared with our ensemble from last year that won the 2014 ImageNet
classification challenge with a 6.66% top-5 error rate).

In this release, we are supplying code and data files containing the trained model parameters for
running the image classifier on:
● Both desktop and mobile environments
● Employing either a C++ or Python API.

In addition, we are providing a tutorial that describes how to use the image recognition system for a
variety of use-cases.
http://www.tensorflow.org/tutorials/image_recognition/index.html
 What do you want in a research system?
● Ease of expression: for lots of crazy ML ideas/algorithms
● Scalability: can run experiments quickly
● Portability: can run on wide variety of platforms
● Reproducibility: easy to share and reproduce research
● Production readiness: go from research to real products
 TensorFlow:
Second Generation Deep Learning System
If we like it, wouldn’t the rest of the world like it, too?

Open sourced single-machine TensorFlow on Monday, Nov. 9th

● Flexible Apache 2.0 open source licensing
● Updates for distributed implementation coming soon

http://tensorflow.org/
 Motivations

DistBelief (1st system):

● Great for scalability, and production training of basic kinds of models
● Not as flexible as we wanted for research purposes

Better understanding of problem space allowed us to

make some dramatic simplifications
TensorFlow: Expressing High-Level ML Computations

● Core in C++

Core TensorFlow Execution System

CPU GPU Android iOS ...

TensorFlow: Expressing High-Level ML Computations

● Core in C++
● Different front ends for specifying/driving the computation
○ Python and C++ today, easy to add more

Core TensorFlow Execution System

CPU GPU Android iOS ...

TensorFlow: Expressing High-Level ML Computations

● Core in C++
● Different front ends for specifying/driving the computation
○ Python and C++ today, easy to add more

C++ front end Python front end ...

Core TensorFlow Execution System

CPU GPU Android iOS ...

 Portable
Automatically runs models on range of platforms:

from phones ...

to single machines (CPU and/or GPUs) …

to distributed systems of many 100s of GPU cards

Computation is a dataflow graph

biases Graph of Nodes, also called Operations or ops.

weights Add Relu

MatMul Xent
examples

labels
 s o rs
Computation is a dataflow graph ten
with

biases Edges are N-dimensional arrays: Tensors

weights Add Relu

MatMul Xent
examples

labels
 t a t e
Computation is a dataflow graph
w ith s

'Biases' is a variable Some ops compute gradients −= updates biases

biases

... Add ... Mul −=

learning rate
 ut e d
Computation is a dataflow graph
is t r i b
d

Device A Device B
biases

Add ... Mul −=

...

learning rate

Devices: Processes, Machines, GPUs, etc

 ut e d
Send and Receive Nodes
is t r i b
d

Device A Device B
biases

Add ... Mul −=

...

learning rate

Devices: Processes, Machines, GPUs, etc

 ut e d
Send and Receive Nodes
is t r i b
d

Device A Device B
biases
Send Recv
Add ... Mul −=
...

learning rate

Devices: Processes, Machines, GPUs, etc

 ut e d
Send and Receive Nodes
is t r i b
d

Device A Device B
biases
Send Recv
Add ... Mul Send Recv −=
... Send Recv
Recv

learning rate Send

Devices: Processes, Machines, GPUs, etc

Send and Receive Implementations
● Different implementations depending on source/dest devices
● e.g. GPUs on same machine: local GPU → GPU copy
● e.g. CPUs on different machines: cross-machine RPC
● e.g. GPUs on different machines: RDMA
Extensible
● Core system defines a number of standard operations
and kernels (device-specific implementations of
operations)
● Easy to define new operators and/or kernels
Session Interface
● Extend: add nodes to computation graph
● Run: execute an arbitrary subgraph
○ optionally feeding in Tensor inputs and retrieving Tensor output

Typically, setup a graph with one or a few Extend calls and

then Run it thousands or millions or times
Single Process Configuration
 Distributed Configuration
RPC

RPC RPC RPC

Feeding and Fetching

Run(input={“b”: ...}, outputs={“f:0”})

Feeding and Fetching

Run(input={“b”: ...}, outputs={“f:0”})

 Example: Power method for Eigenvectors
● Simple 5x5 matrix, compute result, iterated K times
● TensorBoard graph visualization
 Under the hood: Power method
● Operators
● Kernel implementations for different devices
● Run call
● Tensor memory management
 Example: Symbolic differentiation
● f(x) = xT * W * x ; now minimize
● Show df/dx = 2*Wx in graph
 TensorFlow Single Device Performance
Initial measurements done by Soumith Chintala
Benchmark Forward Forward+Backward

AlexNet - cuDNNv3 on Torch (Soumith) 32 ms 96 ms

AlexNet - Neon (Soumith) 32 ms 101 ms

AlexNet - cuDNNv2 on Torch (Soumith) 70 ms 231 ms

AlexNet - cuDNNv2 on TensorFlow 0.5 (Soumith) 96 ms 326 ms

See https://github.com/soumith/convnet-benchmarks/issues/66
Two main factors:
(1) various overheads (nvcc doesn’t like 64-bit tensor indices, etc.)
(2) versions of convolutional libraries being used (cuDNNv2 vs. v3, etc.)
 TensorFlow Single Device Performance
Prong 1: Tackling sources of overhead
Benchmark Forward Forward+Backward

AlexNet - cuDNNv3 on Torch (Soumith) 32 ms 96 ms

AlexNet - Neon (Soumith) 32 ms 101 ms

AlexNet - cuDNNv2 on Torch (Soumith) 70 ms 231 ms

AlexNet - cuDNNv2 on TensorFlow 0.5 (Soumith) 96 ms 326 ms

AlexNet - cuDNNv2 on TensorFlow 0.5 (our machine) 97 ms 336 ms

 TensorFlow Single Device Performance
Prong 1: Tackling sources of overhead
Benchmark Forward Forward+Backward

AlexNet - cuDNNv3 on Torch (Soumith) 32 ms 96 ms

AlexNet - Neon (Soumith) 32 ms 101 ms

AlexNet - cuDNNv2 on Torch (Soumith) 70 ms 231 ms

AlexNet - cuDNNv2 on TensorFlow 0.5 (Soumith) 96 ms 326 ms

AlexNet - cuDNNv2 on TensorFlow 0.5 (our machine) 97 ms 336 ms

AlexNet - cuDNNv2 on TensorFlow 0.6 (our machine: soon) 70 ms (+39%) 230 ms (+31%)
 TensorFlow Single Device Performance
Prong 1: Tackling sources of overhead
Benchmark Forward Forward+Backward

AlexNet - cuDNNv3 on Torch (Soumith) 32 ms 96 ms

AlexNet - Neon (Soumith) 32 ms 101 ms

AlexNet - cuDNNv2 on Torch (Soumith) 70 ms 231 ms

AlexNet - cuDNNv2 on TensorFlow 0.5 (Soumith) 96 ms 326 ms

AlexNet - cuDNNv2 on TensorFlow 0.5 (our machine) 97 ms 336 ms

AlexNet - cuDNNv2 on TensorFlow 0.6 (our machine: soon) 70 ms (+39%) 230 ms (+31%)
 TensorFlow Single Device Performance
TF 0.5 vs. 0.6 release candidate measurements (on our machine w/ Titan-X)

Benchmark Forward Forward+Backward

AlexNet - cuDNNv2 on TensorFlow 0.5 97 ms 336 ms

AlexNet - cuDNNv2 on TensorFlow 0.6 (soon) 70 ms (+27%) 230 ms (+31%)

OxfordNet - cuDNNv2 on TensorFlow 0.5 573 ms 1923 ms

OxfordNet - cuDNNv2 on TensorFlow 0.6 (soon) 338 ms (+41%) 1240 ms (+36%)

Overfeat - cuDNNv2 on TensorFlow 0.5 322 ms 1179 ms

Overfeat - cuDNNv2 on TensorFlow 0.6 (soon) 198 ms (+39%) 832 ms (+29%)

 TensorFlow Single Device Performance

Prong 2: Upgrade to faster core libraries like cuDNN v3

(and/or the upcoming v4)

Won’t make it into 0.6 release later this week, but likely in
next release
Single device performance important, but
….
biggest performance improvements come
from large-scale distributed systems with
model and data parallelism
 Experiment Turnaround Time and Research Productivity
● Minutes, Hours:
○ Interactive research! Instant gratification!
● 1-4 days
○ Tolerable
○ Interactivity replaced by running many experiments in parallel
● 1-4 weeks
○ High value experiments only
○ Progress stalls
● >1 month
○ Don’t even try
 Transition
● How do you do this at scale?
● How does TensorFlow make distributed training easy?
 Model Parallelism
● Best way to decrease training time: decrease the step
time
● Many models have lots of inherent parallelism
● Problem is distributing work so communication doesn’t
kill you
○ local connectivity (as found in CNNs)
○ towers with little or no connectivity between towers (e.g. AlexNet)
○ specialized parts of model active only for some examples
 Exploiting Model Parallelism
On a single core: Instruction parallelism (SIMD). Pretty much
free.

Across cores: thread parallelism. Almost free, unless across

sockets, in which case inter-socket bandwidth matters (QPI on
Intel).

Across devices: for GPUs, often limited by PCIe bandwidth.

Across machines: limited by network bandwidth / latency

Model Parallelism
Model Parallelism
Model Parallelism
 Data Parallelism
● Use multiple model replicas to process different
examples at the same time
○ All collaborate to update model state (parameters) in shared
parameter server(s)
● Speedups depend highly on kind of model
○ Dense models: 10-40X speedup from 50 replicas
○ Sparse models:
■ support many more replicas
■ often can use as many as 1000 replicas
 Data Parallelism
Parameter Servers

Model
Replicas ...
Data
...
 Data Parallelism
Parameter Servers

Model
Replicas ...
Data
...
 Data Parallelism
Parameter Servers

∆p p

Model
Replicas ...
Data
...
 Data Parallelism
Parameter Servers p’ = p + ∆p

∆p p

Model
Replicas ...
Data
...
 Data Parallelism
Parameter Servers p’ = p + ∆p

p’

Model
Replicas ...
Data
...
 Data Parallelism
Parameter Servers

∆p’ p’

Model
Replicas ...
Data
...
 Data Parallelism
Parameter Servers p’’ = p’ + ∆p

∆p’ p’

Model
Replicas ...
Data
...
 Data Parallelism
Parameter Servers p’’ = p’ + ∆p

∆p’ p’

Model
Replicas ...
Data
...
 Data Parallelism Choices
Can do this synchronously:
● N replicas equivalent to an N times larger batch size
● Pro: No gradient staleness
● Con: Less fault tolerant (requires some recovery if any single machine fails)

Can do this asynchronously:

● Pro: Relatively fault tolerant (failure in model replica doesn’t block other
replicas)
● Con: Gradient staleness means each gradient less effective

(Or hybrid: M asynchronous groups of N synchronous replicas)

 Data Parallelism Considerations
Want model computation time to be large relative to time to
send/receive parameters over network
Models with fewer parameters, that reuse each parameter multiple times in the
computation

● Mini-batches of size B reuse parameters B times

Certain model structures reuse each parameter many times within each example:

● Convolutional models tend to reuse hundreds or thousands of times per

example (for different spatial positions)
● Recurrent models (LSTMs, RNNs) tend to reuse tens to hundreds of times
(for unrolling through T time steps during training)
 Success of Data Parallelism
● Data parallelism is really important for many of Google’s
problems (very large datasets, large models):
○ RankBrain uses 500 replicas
○ ImageNet Inception training uses 50 GPUs, ~40X
speedup
○ SmartReply uses 16 replicas, each with multiple GPUs
○ State-of-the-art on LM “One Billion Word” Benchmark
model uses both data and model parallelism on 32
GPUs
10 vs 50 Replica Inception Synchronous Training

50 replicas
10 replicas

Hours
10 vs 50 Replica Inception Synchronous Training

50 replicas
10 replicas
19.6 vs. 80.3 (4.1X)

5.6 vs. 21.8 (3.9X)

Hours
 Using TensorFlow for Parallelism
Trivial to express both model parallelism as well as data
parallelism

● Very minimal changes to single device model code

 Devices and Graph Placement
● Given a graph and set of devices, TensorFlow
implementation must decide which device executes
each node
 Full and Partial Device Constraints (Hints)
Devices are named hierarchically:
/job:localhost/device:cpu:0
/job:worker/task:17/device:gpu:3
/job:parameters/task:4/device:cpu:0

Client can specify full or partial constraints for nodes in

graph:
“Place this node on /job:localhost/device:gpu:2”

“Place this node on /device:gpu:*”

 Placement Algorithm
Given hints, plus a cost model (node execution time
estimates and Tensor size estimates), make placement
decisions

● Current relatively simple greedy algorithm

● Active area of work

Show CIFAR10 placement TensorBoard.

 Example: LSTM [Hochreiter et al, 1997]

● From research paper to code

 Sequence-to-Sequence Model
Target sequence
[Sutskever & Vinyals & Le NIPS 2014] X Y Z Q

A B C D __ X Y Z

Input sequence
Sequence-to-Sequence
● Active area of research
● Many groups actively pursuing RNN/LSTM
○ Montreal
○ Stanford
○ U of Toronto
○ Berkeley
○ Google
○ ...
● Further Improvements
○ Attention
○ NTM / Memory Nets
○ ...
 Sequence-to-Sequence
● Translation: [Kalchbrenner et al., EMNLP 2013][Cho et al., EMLP 2014][Sutskever & Vinyals & Le, NIPS
2014][Luong et al., ACL 2015][Bahdanau et al., ICLR 2015]

● Image captions: [Mao et al., ICLR 2015][Vinyals et al., CVPR 2015][Donahue et al., CVPR 2015][Xu et al.,
ICML 2015]

● Speech: [Chorowsky et al., NIPS DL 2014][Chan et al., arxiv 2015]

● Language Understanding: [Vinyals & Kaiser et al., NIPS 2015][Kiros et al., NIPS 2015]
● Dialogue: [Shang et al., ACL 2015][Sordoni et al., NAACL 2015][Vinyals & Le, ICML DL 2015]
● Video Generation: [Srivastava et al., ICML 2015]
● Algorithms: [Zaremba & Sutskever, arxiv 2014][Vinyals & Fortunato & Jaitly, NIPS 2015][Kaiser &
Sutskever, arxiv 2015][Zaremba et al., arxiv 2015]
How to do Image Captions?

P(English | French)
Image )
 How?
[Vinyals et al., CVPR 2015] A young girl asleep

W __ A young girl
Human: A young girl asleep on
the sofa cuddling a stuffed
bear.

NIC: A close up of a child

holding a stuffed animal.

NIC: A baby is asleep next to a

teddy bear.
 (Recent) Captioning Results
Source: http://mscoco.org/dataset/#leaderboard-cap
Method Meteor CIDEr LSUN LSUN (2)
Google NIC 0.346 (1) 0.946 (1) 0.273 (2) 0.317 (2)
MSR Capt 0.339 (2) 0.937 (2) 0.250 (3) 0.301 (3)
UCLA/Baidu v2 0.325 (5) 0.935 (3) 0.223 (5) 0.252 (7)
MSR 0.331 (4) 0.925 (4) 0.268 (2) 0.322 (2)
MSR Nearest 0.318 (10) 0.916 (5) 0.216 (6) 0.255 (6)
Human 0.335 (3) 0.910 (6) 0.638 (1) 0.675 (1)
UCLA/Baidu v1 0.320 (8) 0.896 (7) 0.190 (9) 0.241 (8)
LRCN Berkeley 0.322 (7) 0.891 (8) 0.246 (4) 0.268 (5)
UofM/Toronto 0.323 (6) 0.878 (9) 0.262 (3) 0.272 (4)
Human: A close up of two
bananas with bottles in the
background.

BestModel: A bunch of bananas

and a bottle of wine.

InitialModel: A close up of a
plate of food on a table.
Human: A view of inside of a car
where a cat is laying down.

BestModel: A cat sitting on top

of a black car.

InitialModel: A dog sitting in

the passenger seat of a car.
Human: A brown dog laying in a
red wicker bed.

BestModel: A small dog is

sitting on a chair.

InitialModel: A large brown dog

laying on top of a couch.
Human: A man outside cooking
with a sub in his hand.

BestModel: A man is holding a

sandwich in his hand.

InitialModel: A man cutting a

cake with a knife.
Human: Someone is using a
small grill to melt his sandwich.

BestModel: A person is cooking

some food on a grill.

InitialModel: A pizza sitting on

top of a white plate.
Human: A woman holding up a
yellow banana to her face.

BestModel: A woman holding a

banana up to her face.

InitialModel: A close up of a
person eating a hot dog.
Human: A blue , yellow and red
train travels across the tracks
near a depot.

BestModel: A blue and yellow

train traveling down train
tracks.

InitialModel: A train that is

sitting on the tracks.
Pointer Networks Teaser
➢ Goal: Mappings where outputs are (sub)sets of inputs
➢ Travelling Salesman Problem

➢ Convex Hulls
Pointer Networks

⇐
1
2
5
6
1
⇒

x1 x2 x3 x4 x5 x6 x1 x6 x5 x2 x1
y1 y2 y3 y4 y5 y6 ⇒ y1 y6 y5 y2 y1

Poster => Wed. 210C #22

 Neural Conversational Models
● Take movie subtitles (~900M words) or IT HelpDesk chats
● Predict the next dialog from history
i got to go .
no .
i get too emotional when i drink .
have another beer . i 've got to get up early .
no , you don 't . sit down .
i get too emotional when i drink .
will you have another beer ?
i 've got to go ! [Vinyals & Le ICML DL Workshop 2015]
why ?
i got to get up early in the morning .
you 're drunk .
and emotional !
you got to go .
 Smart Reply
Google Research Blog
Incoming Email - Nov 2015

Activate
Small Feed- Smart Reply?
Forward yes/no
Neural Network

Generated Replies

Deep Recurrent
Neural Network
 Example: LSTM

for i in range(20):
m, c = LSTMCell(x[i], mprev, cprev)
mprev = m
cprev = c
 Example: Deep LSTM

for i in range(20):
for d in range(4): # d is depth
input = x[i] if d is 0 else m[d-1]
m[d], c[d] = LSTMCell(input, mprev[d], cprev[d])
mprev[d] = m[d]
cprev[d] = c[d]
 Example: Deep LSTM

for i in range(20):
for d in range(4): # d is depth
input = x[i] if d is 0 else m[d-1]
m[d], c[d] = LSTMCell(input, mprev[d], cprev[d])
mprev[d] = m[d]
cprev[d] = c[d]
 Example: Deep LSTM

for i in range(20):
for d in range(4): # d is depth
with tf.device("/gpu:%d" % d):
input = x[i] if d is 0 else m[d-1]
m[d], c[d] = LSTMCell(input, mprev[d], cprev[d])
mprev[d] = m[d]
cprev[d] = c[d]
 A B C D
GPU6

GPU5 A B C D 80k softmax by

1000 dims
This is very big!
GPU4 Split softmax into
4 GPUs

GPU3

GPU2 1000 LSTM cells

2000 dims per
timestep

GPU1
2000 x 4 =
_ 8k dims per
A B C D A B C sentence
_
 A B C D
GPU6

GPU5 A B C D 80k softmax by

1000 dims
This is very big!
GPU4 Split softmax into
4 GPUs

GPU3

GPU2 1000 LSTM cells

2000 dims per
timestep

GPU1
2000 x 4 =
_ 8k dims per
A B C D A B C sentence
_
 A B C D
GPU6

GPU5 A B C D 80k softmax by

1000 dims
This is very big!
GPU4 Split softmax into
4 GPUs

GPU3

GPU2 1000 LSTM cells

2000 dims per
timestep

GPU1
2000 x 4 =
_ 8k dims per
A B C D A B C sentence
_
 A B C D
GPU6

GPU5 A B C D 80k softmax by

1000 dims
This is very big!
GPU4 Split softmax into
4 GPUs

GPU3

GPU2 1000 LSTM cells

2000 dims per
timestep

GPU1
2000 x 4 =
_ 8k dims per
A B C D A B C sentence
_
 A B C D
GPU6

GPU5 A B C D 80k softmax by

1000 dims
This is very big!
GPU4 Split softmax into
4 GPUs

GPU3

GPU2 1000 LSTM cells

2000 dims per
timestep

GPU1
2000 x 4 =
_ 8k dims per
A B C D A B C sentence
_
 A B C D
GPU6

GPU5 A B C D 80k softmax by

1000 dims
This is very big!
GPU4 Split softmax into
4 GPUs

GPU3

GPU2 1000 LSTM cells

2000 dims per
timestep

GPU1
2000 x 4 =
_ 8k dims per
A B C D A B C sentence
_
 A B C D
GPU6

GPU5 A B C D 80k softmax by

1000 dims
This is very big!
GPU4 Split softmax into
4 GPUs

GPU3

GPU2 1000 LSTM cells

2000 dims per
timestep

GPU1
2000 x 4 =
_ 8k dims per
A B C D A B C sentence
_
 A B C D
GPU6

GPU5 A B C D 80k softmax by

1000 dims
This is very big!
GPU4 Split softmax into
4 GPUs

GPU3

GPU2 1000 LSTM cells

2000 dims per
timestep

GPU1
2000 x 4 =
_ 8k dims per
A B C D A B C sentence
_
 A B C D
GPU6

GPU5 A B C D 80k softmax by

1000 dims
This is very big!
GPU4 Split softmax into
4 GPUs

GPU3

GPU2 1000 LSTM cells

2000 dims per
timestep

GPU1
2000 x 4 =
_ 8k dims per
A B C D A B C sentence
_
 A B C D
GPU6

GPU5 A B C D 80k softmax by

1000 dims
This is very big!
GPU4 Split softmax into
4 GPUs

GPU3

GPU2 1000 LSTM cells

2000 dims per
timestep

GPU1
2000 x 4 =
_ 8k dims per
A B C D A B C sentence
_
 A B C D
GPU6

GPU5 A B C D 80k softmax by

1000 dims
This is very big!
GPU4 Split softmax into
4 GPUs

GPU3

GPU2 1000 LSTM cells

2000 dims per
timestep

GPU1
2000 x 4 =
_ 8k dims per
A B C D A B C sentence
_
 A B C D
GPU6

GPU5 A B C D 80k softmax by

1000 dims
This is very big!
GPU4 Split softmax into
4 GPUs

GPU3

GPU2 1000 LSTM cells

2000 dims per
timestep

GPU1
2000 x 4 =
_ 8k dims per
A B C D A B C sentence
_
 TensorFlow Queues

Input prefetching ...

Grouping similar examples

Randomization/Shuffling Dequeue

Enqueue Queue

...
 Example: Deep LSTMs
● Wrinkles
○ Bucket sentences by length using a queue per length
○ Dequeue when a full batch of same length has
accumulated
○ N different graphs for different lengths
○ Alternative: while loop
 Expressing Data Parallelism
# We use the ReplicaDeviceSetter() device function to automatically
# assign Variables to the 'ps' jobs.
with tf.device(“/cpu:0”):
# Create the Mnist model.
model = MnistModel(batch_size=16, hidden_units=200)

# Get an initialized, and possibly recovered session.

sess = tf.Session()

# Train the model.

for local_step in xrange(FLAGS.max_steps):
_, loss, step = sess.run([model.train_op, model.loss, model.global_step])
if local_step % 1000 == 0:
print "step %d: %g" % (step, loss)
 Expressing Data Parallelism
# We use the ReplicaDeviceSetter() device function to automatically
# assign Variables to the 'ps' jobs.
with tf.device(tf.ReplicaDeviceSetter(parameter_devices=10)):
# Create the Mnist model.
model = MnistModel(batch_size=16, hidden_units=200)

# Create a Supervisor. It will take care of initialization, summaries,

# checkpoints, and recovery. When multiple replicas of this program are running,
# the first one, identified by --task=0 is the 'chief' supervisor (e.g., initialization, saving)
supervisor = tf.Supervisor(is_chief=(FLAGS.task == 0), saver=model.saver)

# Get an initialized, and possibly recovered session.

sess = supervisor.PrepareSession(FLAGS.master_job)

# Train the model.

for local_step in xrange(int32_max):
_, loss, step = sess.run([model.train_op, model.loss, model.global_step])
if step >= FLAGS.max_steps:
break
if local_step % 1000 == 0:
print "step %d: %g" % (step, loss)
Asynchronous Training
● Unlike DistBelief, no separate parameter server system:
○ Parameters are now just stateful nodes in the graph
Synchronous Variant
 Network Optimizations
● Neural net training very tolerant of reduced precision
● e.g. drop precision to 16 bits across network

Device A Device B

params Send Recv

Mat
...
Mul
Input
 Network Optimizations
● Neural net training very tolerant of reduced precision
● e.g. drop precision to 16 bits across network

Device A Device B

params ToFP16 Send Recv ToFP32

Mat
...
Mul
Input
 Quantization for Inference
● Need even less precision for inference
● 8-bit fixed point works well, but many ways of
quantizing
● Critical for things like mobile devices
○ w/quantization, high-end smart phone can run
Inception model at >6 frames per second (fps)
 Open Source Status for Distributed TensorFlow
Multi GPU in single machine already in open source release

● See 4-GPU CIFAR10 training example in repository

Distributed implementation coming soon:

● GitHub tracking issue: github.

com/tensorflow/tensorflow/issues/23
 Concluding Remarks
● Model and Data Parallelism enable great ML work:
○ Neural Machine Translation: ~6x speedup on 8 GPUs
○ Inception / Imagenet: ~40x speedup on 50 GPUs
○ RankBrain: ~300X speedup on 500 machines
● A variety of different parallelization schemes are easy to
express in TensorFlow
 Concluding Remarks
● Open Sourcing of TensorFlow
○ Rapid exchange of research ideas (we hope!)
○ Easy deployment of ML systems into products
○ TensorFlow community doing interesting things!
 A Few TensorFlow Community Examples
● DQN: github.com/nivwusquorum/tensorflow-deepq
● NeuralArt: github.com/woodrush/neural-art-tf
● Char RNN: github.com/sherjilozair/char-rnn-tensorflow
● Keras ported to TensorFlow: github.com/fchollet/keras
● Show and Tell: github.com/jazzsaxmafia/show_and_tell.tensorflow
● Mandarin translation: github.com/jikexueyuanwiki/tensorflow-zh
...
github.com/nivwusquorum/tensorflow-deepq
github.com/woodrush/neural-art-tf
github.com/sherjilozair/char-rnn-tensorflow
github.com/fchollet/keras
github.com/jazzsaxmafia/show_and_tell.tensorflow
github.com/jikexueyuanwiki/tensorflow-zh
 Google Brain Residency Program

New one year immersion program in deep learning research

Learn to conduct deep learning research w/experts in our team

● Fixed one-year employment with salary, benefits, ...
● Goal after one year is to have conducted several research projects
● Interesting problems, TensorFlow, and access to computational resources
 Google Brain Residency Program

Who should apply?

● people with BSc, MSc or PhD, ideally in CS, mathematics or statistics
● completed coursework in calculus, linear algebra, and probability, or equiv.
● programming experience
● motivated, hard working, and have a strong interest in deep learning
 Google Brain Residency Program
Program Application & Timeline
DEADLINE: January 15, 2016
 Google Brain Residency Program

For more information:

g.co/brainresidency

Contact us:
[email protected]