Advanced Spark and TensorFlow Meetup May 26, 2016

Advanced and
Meetup
May 26, 2016

Meetup Agenda
Meetup Updates (Chris F)
Technology Updates (Chris F)
Spark + TensorFlow Model Serving/Deployment (Chris F)
Neural Net and TensorFlow Landscape (Chris F)
TensorFlow Core, Best Practices, Hidden Gems (Sam A)
TensorFlow Distributed (Fabrizio M)

Meetup Updates
New Sponsor
Meetup Metrics
Co-presenters
Sam Abrahams
“LA Kid (from DC)
Fabrizio Milo
“Italian Stallion”
Chris Fregly
“Chicago Fire”

advancedspark.com
Workshop - June 4th - Spark ML + TensorFlow

Chris Fregly
Technology Updates
github.com/fluxcapacitor/pipeline
Neural Network Tool Landscape
TensorFlow Tool Landscape
TensorFlow Serving
Spark ML Serving
pipeline.io (Shh… Stealth)

Technology Updates...
Spark 2.0
Kafka 0.10.0 + Confluent 3.0
CUDA 8.0 + cuDNN v5

Whole-Stage Code Gen
● SPARK-12795
● Physically Fuse Together Operators (within a Stage) into 1 Operation
● Avoids Excessive Virtual Function Calls
● Utilize CPU Registers vs. L1, L2, Main Memory
● Loop Unrolling and SIMD Code Generation
● Speeds up CPU-bound workloads (not I/O-bound)
Vectorization (SPARK-12992)
● Operate on Batches of Data
● Reduces Virtual Function Calls
● Fallback if Whole-Stage Not an Option
Spark 2.0: Core

Save/Load Support for All Models and Pipelines!
● Python and Scala
● Saved as Parquet ONLY
Local Linear Algebra Library
● SPARK-13944, SPARK-14615
● Drop-in Replacement for Distributed Linear Algebra library
● Opens up door to my new pipeline.io prediction layer!
Spark 2.0: ML

Kafka v0.10 + Confluent v3.0 Platform
Kafka Streams
New Event-Creation Timestamp
Rack Awareness (THX NFLX!)
Kerberos + SASL Improvements
Ability to Pause a Connector (ie. Maintenance)
New max.poll.records to limit messages retrieved

CUDA Deep Neural Network (cuDNN) v5 Updates
LSTM (Long Short-Term Memory) RNN Support for NLP use cases (6x speedup)
Optimized for NVIDIA Pascal GPU Architecture including FP16 (low precision)
Highly-optimized networks with 3x3 convolutions (GoogleNet)

github.com/fluxcapacitor/pipeline Updates
Tools and Examples
● JupyterHub and Python 3 Support
● Spark-Redis Connector
● Theano
● Keras (TensorFlow + Theano Support)
Code
● Spark ML DecisionTree Code Generator (Janino JVM ByteCode Generator)
● Hot-swappable ML Model Watcher (similar to TensorFlow Serving)
● Eigenface-based Image Recommendations
● Streaming Matrix Factorization w/ Kafka
● Netflix Hystrix-based Circuit Breaker - Prediction Service @ Scale

Interesting Neural Net Use Case

TensorFlow Landscape
TensorFlow Core
TensorFlow Distributed
TensorFlow Serving (similar to Prediction.IO, Pipeline.IO)
TensorBoard (Visualize Neural Network Training)
Playground (Toy)
SkFlow (Scikit-Learn + TensorFlow)
Keras (High-level API for both TensorFlow and Theano)
Models (Parsey McParseface/SyntaxNet)

TensorFlow Serving (Model Deployment)
Dynamic Model Loader
Model Versioning & Rollback
Written in C/C++
Extend to Serve any Model!
Demos!!

Spark ML Serving (Model Deployment)
Same thing except for Spark...
Keep an eye on pipeline.io!

Sam Abrahams
TensorFlow Core
Best Practices
Hidden Gems

TensorFlow Core Terminology
gRPC (?) - Should this go here or in distributed?
TensorBoard Overview (?)
Explaining this will go a long way ->

Who dis?
•My name is Sam Abrahams
•Los Angeles based machine learning engineer
•TensorFlow White Paper Notes
•TensorFlow for Raspberry Pi
•Contributor to the TensorFlow project
samjabrahams @sabraha

This Talk: Sprint Through:
• Core TensorFlow API and terminology
• TensorFlow workflow
• Example TensorFlow Code
• TensorBoard

TensorFlow Programming Model
• Very similar to Theano
• The primary user-facing API is Python, though there is a
partial C++ API for executing models
• Computational code is written in C++, implemented for
CPUs, GPUs, or both
• Integrates tightly with NumPy

TensorFlow Programming
Generally boils down to two steps:
1. Build the computational graph(s) that you’d
like to run
2. Use a TensorFlow Session to run your graph
one or more times

Graph
• The primary structure of a TensorFlow model
• Generally there is one graph per program, but
TensorFlow can support multiple Graphs
• Nodes represent computations or data transformations
• Edges represent data transfer or computational control

What is a data flow graph?
• Also known as a “computational graph”, or just “graph”
• Nice way to visualize series of mathematical computations
• Here’s a simple graph showing the addition of two variables:
a ba
a + b

Components of a Graph
Graphs are composed of two types of elements:
• Nodes
• These are the elliptical shapes in the graph, and represent some form of
computation
• Edges
• These are the arrows in the graph, and represent data flowing from one
node into another node.
NODE NODEEDGE

Why Use Graphs?
• They are highly compositional
• Useful for calculating derivatives
• It’s easier to implement distributed computation
• Computation is already segmented nicely
• Neural networks are already implemented as computational graphs!

Tensors
• N-Dimensional Matrices
• 0-Dimensional → Scalar
• 1-Dimensional → Vector
• 2-Dimensional → Matrix
• All data that moves through a TensorFlow graph is a Tensor
• TensorFlow can convert Python native types or NumPy arrays
into Tensor objects

Tensors
•Python
# Vector
> [1, 2, 3]
# 3-D Tensor
> [[ [0,0], [0,1], [0,2] ],
[ [1,0], [1,1], [1,2] ],
[ [2,0], [2,1], [2,2] ]
# Matrix as NumPy Array
> np.array([[1, 2, 3],
[4, 5, 6],
[7, 8, 9]],
dtype=np.int64)
# Scalar
> 3
# Matrix
> [[1, 2, 3],
[4, 5, 6],
[7, 8, 9]]
•NumPy

Tensors
• Best practice is to use NumPy arrays when directly
defining Tensors
• Can explicitly set data type
• This presentation does not create Tensors with NumPy
• Space is limited
• I am lazy
• Tensors returned by TensorFlow graphs are NumPy
arrays

Operations
• “Op” for short
• Represent any sort of computation
• Take in zero or more Tensors as input, and output zero
or more Tensors
• Numerous uses: perform matrix algebra, initialize
variables, print info to the console, etc.
• Do not run when defined: they must be called from a
TensorFlow Session (coming up later)

Operations
•Quick Example
> import tensorflow as tf
> a = tf.mul(3,5) # Returns handle to new tf.mul node
> sess = tf.Session() # Creates TF Session
> sess.run(a) # Actually runs the Operation
out: 15

Placeholders
• Define “input” nodes
• Specifies information that be provided when graph is run
• Typically used for training data
• Define the tensor shape and data type when created:
import tensorflow as tf
# Create a placeholder of size 100x400
# With 32-bit floating point data type
my_placeholder = tf.placeholder(tf.float32, shape=(100,400))

TensorFlow Session
• In charge of coordinating graph execution
• Most important method is run()
• This is what actually runs the Graph
• It takes in two parameters, ‘fetches’ and ‘feed_dict’
• ‘fetches’ is a list of objects you’d like to get the results for, such
as the final layer in a neural network
• ‘feed_dict’ is a dictionary that maps tensors (often
Placeholders) to values those tensors should use

TensorFlow Variables
• Contain tensor data that persists each time you run your
Graph
• Typically used to hold weights and biases of a machine
learning model
• The final values of the weights and biases, along with
the Graph shape, define a trained model
• Before being run in a Graph, must be initialized (will
discuss this at the Session slide)

TensorFlow Variables
•Old school: Define with tensorflow.Variable()
•Best practices: tf.get_variable()
•Update its information with the assign() method
Import tensorflow as tf
# Create a variable with value 0 and name ‘my_variable’
my_var = tf.get_variable(0, name=‘my_variable’)
# Increment the variable by one
my_var.assign(my_var + 1)

Building the graph!
import tensorflow as tf
# create a placeholder for inputting floating point data
x = tf.placeholder(tf.float32)
# Make a Variable with the starting value of 0
start = tf.Variable(0.0)
# Create a node that is the value of (start + x)
y = start.assign(start + x)
a startx
y = x + start

Running a Session (using previous graph)
• Start a Session with tensorflow.Session()
• Close it when you’re done!
# Open up a TensorFlow Session
# and assign it to the handle 'sess'
sess = tf.Session()
# Important: initialize the Variable
init = tf.initialize_all_variables
sess.run(init)
# Run the graph to get the value of y
# Feed in different values of x each time
print(sess.run(y, feed_dict={x:1})) # Prints 1.0
print(sess.run(y, feed_dict={x:0.5})) # Prints 1.5
print(sess.run(y, feed_dict={x:2.2})) # Prints 3.7
# Close the Session
sess.close()

Devices
• A single device is a CPU, GPU, or other computational
unit (TPU?!)
• One machine can have multiple devices
• “Distributed” → Multiple machines
• Multi-device != Distributed

TensorBoard
• One of the most useful (and underutilized) aspects of
TensorFlow - Takes in serialized information from graph to
visualize data.
• Complete control over what data is stored
Let’s do a brief live demo. Hopefully nothing blows up!

TensorFlow Codebase
Standard and third party C++ libraries (Eigen)
TensorFlow core framework
C++ kernel implementation
Swig
Python client API

TensorFlow Codebase Structure: Core
tensorflow/core : Primary C++ implementations and runtimes
•core/ops – Registration of Operation signatures
•core/kernels – Operation implementations
•core/framework – Fundamental TensorFlow classes and functions
•core/platform – Abstractions for operating system platforms
Based on information from Eugene Brevdo, Googler and TensorFlow member

TensorFlow Codebase Structure: Python
tensorflow/python : Python implementations, wrappers, API
•python/ops – Code that is accessed through the Python API
•python/kernel_tests – Unit tests (which means examples)
•python/framework – TensorFlow fundamental units
•python/platform – Abstracts away system-level Python calls
contrib/ : contributed or non-fully adopted code
Based on information from Eugene Brevdo, Googler and TensorFlow member

Learning TensorFlow: Beyond the Tutorials
● There is a ton of excellent documentation in the code itself-
especially in the C++ implementations
● If you ever want to write a custom Operation or class, you need
to immerse yourself
● Easiest way to dive in: look at your #include statements!
● Use existing code as reference material
● If you see something new, learn something new

Tensor - tensorflow/core/framework/tensor.h
• dtype() - returns data type
• shape() - returns TensorShape representing tensor’s shape
• dims() – returns the number of dimensions in the tensor
• dim_size(int d) - returns size of specified dimension
• NumElements() - returns number of elements
• isSameSize(const Tensor& b) – do these Tensors dimensions match?
• TotalBytes() - estimated memory usage
• CopyFrom() – Copy another tensor and share its memory
• ...plus many more

Some files worth checking out:
tensorflow/core/framework
● tensor_util.h: deep copying, concatenation, splitting
● resource_mgr.h: Resource manager for mutable Operation state
● register_types.h: Useful helper macros for registering Operations
● kernel_def_builder.h: Full documentation for kernel definitions
tensorflow/core/util
● cuda_kernel_helper.h: Helper functions for cuda kernel implementations
● sparse/sparse_tensor.h: Documentation for C++ SparseTensor class

General advice for GPU implementation:
1. If possible, always register for GPU, even if it’s not a full implementation
○ Want to be able to run code on GPU if it won’t bottleneck the system
2. Before writing a custom GPU implementation, sanity check! Will it help?
○ Not everything benefits from parallelization
3. Utilize Eigen!
○ TensorFlow’s core Tensor class is based on Eigen
4. Be careful with memory: you are in charge of mutable data structures

Fabrizio Milo
TensorFlow Distributed
githbu.com/Mistobaan twitter.com/fabmilo

gRPC (http://www.grpc.io/)
- A high performance, open source, general RPC framework that puts mobile and HTTP/2
first.
- Protocol Buffers
- HTTP2
HTTP2
HTTP2

gRPC
python.Tensor python.Tensor
tensor.proto

Distributed Tensorflow - Terminology
- Cluster
- Jobs
- Tasks

Distributed Tensorflow - Terminology
- Cluster: ClusterSpec
- Jobs: Parameter Server, Worker
- Tasks: Usually 0, 1, 2, …

Example of a Cluster Spec
{
"ps":[
"8.9.15.20:2222",
],
"workers":[
"8.34.25.90:2222",
"30.21.18.24:2222",
"4.17.19.14:2222"
]
}

One Session One Server One Job One Task

One Session One Server One Job One Task
ServerSession

Use Cases
- Training
- Hyper Parameters Optimization
- Ensembling

tf.python.training.session_manager.SessionManager
1. Checkpointing trained variables as the training
progresses.
2. Initializing variables on startup, restoring them from the
most recent checkpoint after a crash, or wait for
checkpoints to become available.

tf.python.training.supervisor.Supervisor
# Single Program
with tf.Graph().as_default():
...add operations to the graph...
# Create a Supervisor that will checkpoint the model in '/tmp/mydir'.
sv = Supervisor(logdir='/tmp/mydir')
# Get a Tensorflow session managed by the supervisor.
with sv.managed_session(FLAGS.master) as sess:
# Use the session to train the graph.
while not sv.should_stop():
sess.run(<my_train_op>)

tf.python.training.supervisor.Supervisor
# Multiple Replica Program
is_chief = (server_def.task_index == 0)
server = tf.train.Server(server_def)
with tf.Graph().as_default():
...add operations to the graph...
# Create a Supervisor that uses log directory on a shared file system.
# Indicate if you are the 'chief'
sv = Supervisor(logdir='/shared_directory/...', is_chief=is_chief)
# Get a Session in a TensorFlow server on the cluster.
with sv.managed_session(server.target) as sess:
# Use the session to train the graph.
while not sv.should_stop():
sess.run(<my_train_op>)

Use Cases: Asynchronous Training

Use Cases: Synchronous Training
tf.train.SyncReplicasOptimizer
This optimizer avoids stale gradients by collecting gradients from all replicas,
summing them, then applying them to the variables in one shot, after
which replicas can fetch the new variables and continue.

tf.train.replica_device_setter
The device setter will automatically place Variables ops on separate
parameter servers (ps). The non-Variable ops will be placed on the workers.
tf.train.replica_device_setter(cluster_def)
with tf.device(device_setter):
pass

Training Replication In Graph Between Graphs
Asynchronous
Synchronous
Use Cases: Training

/job:ps /job:worker
In-Graph Replication
/task:0 /task:0
/task:1/task:1
with tf.device("/job:ps/task:0"):
weights_1 = tf.Variable(...)
biases_1 = tf.Variable(...)
with tf.device("/job:ps/task:1"):
weights_2 = tf.Variable(...)
biases_2 = tf.Variable(...)
with tf.device("/job:worker/task:0"):
input, labels = ...
layer_1 = tf.nn.relu(...)
with tf.device("/job:worker/task:0"):
train_op = ...
logits = tf.nn.relu(...)
with tf.Session() as sess:
for _ in range(10000):
sess.run(train_op)
Client

# To build a cluster with two ps jobs on hosts ps0 and ps1, and 3
worker
# jobs on hosts worker0, worker1 and worker2.
cluster_spec = {
"ps": ["ps0:2222", "ps1:2222"],
"worker": ["worker0:2222", "worker1:2222", "worker2:2222"]}
with tf.device(tf.replica_device_setter(cluster=cluster_spec)):
# Build your graph
v1 = tf.Variable(...) # assigned to /job:ps/task:0
# Run compute
...
/job:ps /job:worker
Replication Between Graph
/task:0 /task:0
/task:1/task:1

Use Cases: Train HyperParameter Optimization
- Grid Search
- Random Search
- Gradient Optimization
- Bayesian

Ensemble
Ensembling
Model 0
Model 1
Model 2
Ensemble

Comparison with other frameworks
mxnet

Surprise Demo : Rasberry Pi Cluster
Distributed TensorFlow Cluster
Distributed TensorFlow
TensorBoard Visualizations

Workshop Demo - LARGE GCE Cluster
Super-Large Cluster
Attendee GCE (Google Cloud Engine) Spec
50 GB RAM, 100 GB SSD, 8 CPUs (No GPUs)
TODO:
Build Script for Each Attendee to Run as Either a Worker or Parameter Server
Figure out split between Worker and Parameter Server
ImageNet
TODO: Train full 64-bit precision, quantitize down to 8-bit (Sam A)

Workshop Demo - Initial Cluster
8.34.215.90 (Sam)
130.211.128.240 (Fabrizio)
104.197.159.134 (Fregly)
Cluster Spec: “8.34.215.90, 130.211.128.240, 104.197.159.134”
SSH PEM file: http://advancedspark.com/keys/pipeline-training-gce.pem
chmod 600 pipeline-training-gce.pem
Username: pipeline-training
Password: password9
sudo docker images (VERIFY fluxcapacitor/pipeline)
START: sudo docker run -it --privileged --name pipeline --net=host -m 48g fluxcapacitor/pipeline bash

TensorFlow Operations: Hidden Gems
- Go over macros, functions provided in the library, but not mentioned in
documentation
- Brief brief overview of writing custom op
- Testing GPU code
- Leveraging Eigen and existing code
- Python wrapper

Advanced Spark and TensorFlow Meetup May 26, 2016

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Advanced Spark and TensorFlow Meetup May 26, 2016 (20)

More from Chris Fregly (20)

Recently uploaded (20)

Advanced Spark and TensorFlow Meetup May 26, 2016