251

CONTENTS
GETTING STARTED WITH

TensorFlow
 Introduction to TensorFlow
 Three Layers of TensorFlow Architecture
 Quick Python Example
 Use Cases
BY TIMOTHY SPANN  Setup and Installation...and more!

INTRODUCTION TO TENSORFLOW programs are usually structured into a construction phase, which
assembles a data graph, and an execution phase, which uses a
TensorFlow is a deep learning library from Google that is open-
session to execute operations in the graph.
source and available on GitHub. TensorFlow excels at numerical
computing, which is critical for deep learning. It has a rich set of Questioning whether TensorFlow is ready for production
application programming interfaces in most major languages and environments belies the fact that TensorFlow has been in the
environments needed for deep learning projects: Python, C, C++, open-source community for almost two years and has been
Rust, Haskell, Go, Java, Android, IoS, Mac OS, Windows, Linux, and used by many companies. TensorFlow excels in production
Raspberry Pi. The primary unit in TensorFlow is a tensor. A tensor environments due to its distributed and parallel design and its
consists of a set of primitive values shaped into an array of any ability to access data from Hadoop HDFS. Another key feature of
number of dimensions. These massive numbers of large arrays are TensorFlow is TensorFlow Serving, which is a high-performance
the reason that GPUs and other processors designed to do floating production server for deploy and serve machine learning models
point mathematics excel at speeding up these algorithms. using gRPC as its network protocol.

TENSORS THREE LAYERS OF TENSORFLOW ARCHITECTURE

A tensor is a set of primitives in a multidimensional array, ranked
by number of dimensions 1. High-level libraries: Python, TensorBoard, Java, and more.
2. TensorFlow core: Neural net ops, graph execution engine,
0: 5.0
1: [5.0, 10.0, 15.0, 20.0, 25.0} and more.
2: [[5.0, 10.0], [15.0, 20.0], [25.0, 30.0], [35.0,
40.0, 45.0]] 3. Platforms: CPUs, GPUs, TPUs, Android, and iOS across all
major operating systems.

In TensorFlow programs, you will need to use variables to

QUICK PYTHON EXAMPLE
hold tensors in memory. Some tensors are constants and not
variables; for example, static numbers. An important note is that
before you start using variables, you must initialize them to a import tensorflow as tf
value. TensorFlow first builds a graph of all the operation to be five = tf.constant(“00”)
tf.Variable(0, name=”variablename”)
done, and then when a “session” is called, it “runs” the graph. It’s session = tf.Session()
built to be scalable, by changing internal data representation to
tensors (AKA multi-dimensional arrays). file_writer = tf.summary.FileWriter(‘tflogs’, session.
graph)

DOING SOME SIMPLE MATH run_options = tf.RunOptions(trace_level=tf.RunOptions.

FULL_TRACE)
run_metadata = tf.RunMetadata()
import tensorflow as tf

summary, result = session.run(fetches=five, feed_dict=None,

firstnumber = tf.constant([10, 20, 30, 35, 40, 50])
options=run_options, run_metadata=run_metadata)
secondnumber = tf.Variable(firstnumber + 50)
print(‘Value %s’ % result)
file_writer.add_run_metadata(run_metadata, ‘result %s’ %
with tf.Session() as session:
result)
session.run(tf.global_variables_initializer())
session.close()
print(session.run(secondnumber))

Neural networks have been around for a long time, but now they
The APIs are well-documented and well-designed, so it was easy,
can be run in a reasonable amount of time due to more powerful
for example, for me to port the LabelImage Java example to work
computers, GPUs, and elastic cloud computing. Google has made
as a processor for Apache NiFi. Some people prefer a higher-level
neural networks mainstream through the public availability of
interface for neural network programming, such as Keras. Keras is
TensorFlow on all major clouds and platforms to be used by
available for Python and works not only for TensorFlow but also
researchers, data scientists, data engineers, cloud developers,
for CNTK and Theano. I recommend using Python 3.7 and starting
mobile developers, and web application users.
with plain TensorFlow before you investigate Keras. TensorFlow

DZONE.COM | © DZONE, INC. VISIT DZONE.COM/REFCARDZ FOR MORE!

 2 GETTING STARTED WITH TENSORFLOW

An example implementation of a neural network would be to environment, but I like with keep it simple to Python 3.6 and
define an architecture, load data to the model, break up data into pip3. For some environments, you may need to be root or
batches to preprocess, convert it, and use it for training. As the have sudo access.
model gets trained in increments, the model is saved for reuse.
apt-get install python3-pip python3-dev
Future runs test the model on new data to check performance.
pip3 install tensorflow
After that, you continue to train over many iterations and as much pip3 install grpcio==1.4.0
training data as you can obtain. This is helped by fast networks,
fast CPUs, fast GPUs, large RAM, large hard drives, and generally
For GPU
fast computers.
pip3 install tensorflow-gpu
USE CASES

• Speech recognition: Recognize what people are saying and Sometimes, that doesn’t work. Look at GitHub or try:
convert it to text.
sudo pip3 install --upgrade latesturlforyourarchitecture
• Image recognition: Describe, in text, what an image is. from github.

• Sentiment analysis: Determine if a human sentence is

positive, negative, or neutral. I also recommend you have the TensorFlow source, models, and
TensorBoard GitHub repositories downloaded. You should also
• Text summarization: Summarize a longer article automatically.
install JDK 8+, Bazel, and all build tools required for your operating
• Mobile image and video processing: Process images and system. I recommend you only work with TensorFlow on Ubuntu
videos from a mobile device. or OSX. If you are interested in running on Raspberry PI, NVidia
TX1, or other devices, make sure you follow specific guidelines for
• Language translation: Translate between human languages. those builds. They will often have specialized instructions, utilities,
• Object detection in photos: Determine what items are in versions, or even forks of source code to utilize.
photos and box them.
On OSX, you should have homebrew installed:
• Image captioning: Add captions and possible paragraphs
describing an image. brew upgrade bazel

• Chatbot: Intelligence response and routing of live people

with TensorFlow-generated replies. With older versions, or if something goes wrong, you may need to
try some other installation options.
• Cancer detection: Compare facial and body photos with
sudo easy_install --upgrade six
photos of known cancer patients.
sudo pip install --upgrade tensorflow
pip install grpc

SETUP AND INSTALLATION

Serious data scientists will want a powerful desktop computer I recommend checking that you have TensorFlow installed by
that has 128GB of RAM or more, several NVidia high-end CUDA running pip show and then running our example program above.
cards, fast SSD storage, and a recent version of Ubuntu. The main pip show tensorflow
platform to run TensorFlow is Ubuntu-based Linux. If you have an Name: tensorflow
NVidia GPU that supports CUDA, you need to install those NVidia Version: 1.3.0
Summary: TensorFlow helps the tensors flow
libraries to make sure you take advantage of all the GPUs on the
Home-page: http://tensorflow.org/
device. This is critical for performance, as the large number of Author: Google Inc.
matrix multiplication requires floating point capable processors Author-email: [email protected]
like GPUS. If you do not have a GPU available, you will face License: Apache 2.0
difficulties in running data that’s not pre-trained. Even a decent Location: /usr/local/lib/python2.7/site-packages
Requires: wheel, backports.weakref, protobuf, numpy,
machine will stall your progress if it is not equipped with a CUDA-
mock, tensorflow-tensorboard, six
capable, many-core GPU. On several occasions, I have attempted
to train IM2TXT a PowerBook and gave up after days of running
with little progress. For those really interested in the cutting edge or facing issues,
you can try the current nightly build for your platform here.
If you are prepared to use pre-trained models or just try out a
basic example, feel free to other a non-GPU enabled device.
GETTING STARTED
Whatever platform you have picked, TensorFlow has a lot of Most people start with MNIST, which is a dataset of handwritten
means of installation: source build from GitHub, Python virtual digits. This is a fun, well-documented example to start with, and
env, Python pip, Docker, and Anaconda. Anaconda is a nice is often thought of as the “Hello World” example. Before you start

DZONE.COM | © DZONE, INC.

 3 GETTING STARTED WITH TENSORFLOW

with this, you must know Python and some basics on running • tf.Variable
shell programs in your environment. Creates a variable from a tensor.

As we documented in our simple example, start off by importing • tf.decode_csv

TensorFlow. From there, you can create constants and variables, Converts CSV files to tensors, with each column mapping to a
create a session, and then run the session. A session is needed to tensor.
execute your graph(s).
• tf.decode_base64
In the Java version, the key imports are: Decodes base-64-encoded tensor strings.

import org.tensorflow.DataType; • tf.subtract

import org.tensorflow.Graph;
import org.tensorflow.Output;
Subtracts one tensor from another (must be the same type).
import org.tensorflow.Session;
import org.tensorflow.Tensor; • tf.add
Adds two tensors of the same type together, for numbers and
strings.
MAVEN BUILD FOR JAVA

<dependency> • tf.to_int32
<groupId>org.tensorflow</groupId> Converts tensor primitives to int32.
<artifactId>tensorflow</artifactId>
<version>1.3.0</version> • tf.multiply
</dependency>
Multiplies two tensors of the same type together, for numbers.

In the Java version, you start with a graph, import a graph • tf.div
definition, create a session, and return a tensor. I highly Divides numerator by denominator of real numeric types and
recommend looking at the Java and Android examples provided return quotient.
in the TensorFlow GitHub repository.
• tf.abs
The absolute value of the tensor.
EXAMPLE RUN
• tf.negative
import tensorflow as tf Returns negative value of numeric tensor.
firstnumber = tf.constant([10, 20, 30, 35, 40, 50])
secondnumber = tf.Variable(firstnumber + 50)
• tf.maximum
Returns the maximum of two tensors for a subset of numeric types.
with tf.Session() as session:
session.run(tf.global_variables_initializer()) • tf.minimum
print(session.run(secondnumber))
Returns the minimum of two tensor for a subset of numeric types.
python3 testtf2.py
2017-10-02 14:40:33.009719: W tensorflow/core/platform/ • tf.string_to_number
cpu_feature_guard.cc:45] The TensorFlow library wasn’t Converts a string tensor to a specified numeric type; specify
compiled to use SSE4.2 instructions, but these are out_type=tf.DType, where DType is a subset of numeric types.
available on your machine and could speed up CPU
computations.
2017-10-02 14:40:33.009742: W tensorflow/core/platform/
• tf.concat
cpu_feature_guard.cc:45] The TensorFlow library wasn’t Concatenates a tensor, along one specified dimension.
compiled to use AVX instructions, but these are available
on your machine and could speed up CPU computations. • tf.fill
2017-10-02 14:40:33.009747: W tensorflow/core/platform/
Creates a tensor filled with a specified scalar value.
cpu_feature_guard.cc:45] The TensorFlow library
wasn’t compiled to use AVX2 instructions, but these
are available on your machine and could speed up CPU • tf.tuple
computations. Groups tensors together.
2017-10-02 14:40:33.009751: W tensorflow/core/platform/
cpu_feature_guard.cc:45] The TensorFlow library wasn’t • tf.zeros
compiled to use FMA instructions, but these are available
on your machine and could speed up CPU computations.
Creates a tensor populated with zeros of a certain numeric
[ 60 70 80 85 90 100] type shape.

• tf. convert_to_tensor
COMMON TENSORFLOW OPERATIONS Converts the value to a tensor of specified type.

• tf.rank • tf.while_loop
Returns the rank of a tensor.
The while loop.
• tf.constant
• tf.case
Creates a constant from a value, type, and shape. The case operations on bools.

DZONE.COM | © DZONE, INC.

 4 GETTING STARTED WITH TENSORFLOW

• tf.count_up_to • tf.image.resize_images
Increments a mutable tensor of int32 or int64 to an int limit. Resizes images to sizes using one of four predefined methods.

• tf.Print • tf.image.rot90
Prints out a debug message of tensors and messages. Rotates an image counter-clockwise 90 degrees.

• tf.is_nan • tf.image.transpose_image
Returns elements of tensors that are not a number. Transposes an image by swapping first two dimensions.

• tf.is_finite • tf.image.adjust_brightness
Returns elements of tensors that are finite, for half, float32, Adjusts image brightness.
and float64. • tf.image.adjust_contrast
Adjusts contrast of images.
• tf.logical_and
Returns truth value of each boolean element of two tensors. • tf.image.adjust_hue
Adjusts hue of images.
• tf.logical_not
Returns truth value of not each boolean element of a tensor. • tf.image.adjust_gamma
Runs gamma correction on input image.
• tf.logical_or
Returns truth value of each boolean element of two tensors. • tf.image.adjust_saturation
Adjusts saturation channel in image.
• tf.logical_xor
Returns truth value of each of boolean elements, exclusive or. • tf.image.flip_left_right
Flips an image horizontally.
• tf.equal
Returns truth value of first tensor == second tensor, element- • tf.image.flip_up_down
wise, on two tensors. Flips an image vertically.

• tf.not_equal • tf.image.draw_bounding_boxes
Returns truth value of first tensor != second tensor, element- Draws bounding boxes on a batch of images.
wise, on two tensors. • tf.nn
A large collection of neural network ops.
• tf.greater_equal
Returns truth value of first tensor >= second tensor, element- • tf.contrib
wise, on two tensors. Contributed libraries around keras, audio, video, TFLearn, and more.

• tf.greater • tf.Session()
Returns truth value of first tensor > second tensor, element- An instance of a session that is the environment to execute
wise, on two tensors. operations in a graph to compute tensors.

• tf.less_equal
TRAINING
Returns truth value of first tensor <= second tensor, element-
wise, on two tensors. As we mentioned in an earlier section, you should not try your own
training unless your problem space is small enough or you have
• tf.less sufficient hardware to run your training. Training examples like
Returns truth value of first tensor < second tensor, element- CIFAR-10 will also require downloading anywhere from hundreds
wise, on two tensors. of megabytes of data to terabytes of training data, especially
when dealing with images and videos. You will be working on
• tf.where
datasets of thousands of items and thousands of steps. These
Returns elements from either of two tensors based on condition.
numbers can increase astronomically depending on your problem
• tf.image.decode_jpeg set and data. A cool dataset to check out is SVHN, which contains
Decodes a JPEG image to a uint8 tensor. real-world street house numbers.

• tf.image.decode_png IMAGE RECOGNITION WITH INCEPTION V3

Decodes a PNG image to a uint8 or uint16 tensor.
A very popular application of TensorFlow is image recognition.
• tf.image.decode_gif Whether it’s from a Mac, a Raspberry Pi, or a mobile phone,
Decodes a GIF image to a uint8 tensor. this is a great use of TensorFlow. If you use Inception V3, you
will get a pretty good set of data, which will result in good
• tf.image_decode_bmp results. One thing to note is that you want a clear picture without
Decodes a BMP image to a uint8 tensor.

DZONE.COM | © DZONE, INC.

 5 GETTING STARTED WITH TENSORFLOW

too much noise in it in order for recognition to have a high USING TENSORBOARD FOR VISUALIZATION AND DEBUGGING
matching percentage. You will need to clone the TensorBoard GitHub repo. Then, you
can run TensorBoard pointing to a recent log. We have generated
For image recognition with the Inception v3 application, I one from the example given in this Refcard.
recommend you download the pre-trained model here.
bazel run tensorboard -- -- logdir /opt/demo/tflogs
The example has a pre-trained 80-megabyte model from the
TensorFlow models repo, which lets you quickly classify images
against that model. This will run on a smartphone, a Raspberry Pi, The summary data is logged and sent to the board for
or your laptop easily in a few seconds. visualization. It is a great tool for learning what is going on and
learning how to improve your results as you are training and
EXAMPLE tweaking your algorithms.

python classify_image.py --image_file /opt/demo/ SUMMARY

SolarPanelsOnRoof.jpg
solar dish, solar collector, solar furnace (score = TensorFlow is a really useful library with many practical uses. The
0.98316)
Python binding is very mature and usable. When you decide to
window screen (score = 0.00196)
manhole cover (score = 0.00070) run in production consider using TensorFlow Serving, TensorFlow
radiator (score = 0.00041) on Hadoop, or TensorFlow on Spark.
doormat, welcome mat (score = 0.00041)
If you are just beginning to think about deep learning, neural
networks, artificial intelligence, or just something practical like
image analysis, then evaluate TensorFlow for your needs. I highly
recommend you start with the basics and make sure you have a
stable Python 3.7 environment set up with all the modules that
you will need and that you have basic Python applications that
run at a reasonable rate. At a minimum, you will need NumPy
installed on your TensorFlow machine.

Step one is to get everything installed. Then, try the basic

example we have included in this Refcard. The next step is to
determine which of your use cases matches the ones I have listed.
If you don’t have one matching, search GitHub for TensorFlow
projects. If you find nothing, perhaps it’s time to look at other
machine learning projects or other use cases to try.

A B O U T T H E AU T H O R
TIMOTHY SPANN is a Big Data Solutions Engineer. He helps educate and disseminate performant
open source solutions for Big Data initiatives to customers and the community. With over 15 years of
experience in various technical leadership, architecture, sales engineering, and development roles, he
is well-experienced in all facets of Big Data, cloud, IoT, and microservices. As part of his community
efforts, he also runs the Future of Data Meetup in Princeton.

DZone communities deliver over 6 million pages each month to more than 3.3 million
software developers, architects and decision makers. DZone offers something for
everyone, including news, tutorials, cheat sheets, research guides, feature articles, DZONE, INC. REFCARDZ FEEDBACK
150 PRESTON EXECUTIVE DR. WELCOME
source code and more. "DZone is a developer's dream," says PC Magazine.
[email protected]
CARY, NC 27513

Copyright © 2017 DZone, Inc. All rights reserved. No part of this publication may be reproduced, stored in a retrieval SPONSORSHIP
system, or transmitted, in any form or by means electronic, mechanical, photocopying, or otherwise, without prior 888.678.0399 OPPORTUNITIES
written permission of the publisher. 919.678.0300 [email protected]