Introduction to TensorFlow

by Aadil Hayat

Deep Learning

Deep Learning

• Deep Learning is a new area of Machine Learning research, which has been introduced with the objective of moving Machine Learning closer to one of its original goals: Artificial Intelligence�
• Deep Learning allows computer to build complex concepts out of simpler concepts.�
• It is a part of broader family of Machine Learning methods based on learning representations of data.�
• The typical example of a deep learning model is the feedforward neural network or multilayer perceptron (MLP)

Deep Learning

Deep Learning

Applications

• Speech Recognition�
• Machine Translations�
• Natural Language Processing�
• Object Recognition�
• Bioinformatics

TensorFlow

What is it?

• Open Source�
• Numerical computation library�
• Using Data flow graphs�
• Flexible Architecture: CPUs, GPUs, desktop, server or mobile�
• Available in Python, C++, Java and Go�
• Developed by Google Brain Team for research��

Why Tensorflow?

Credits: Benchmarking State-of-the-Art Deep Learning Software Tools. Shi et al. Feb 2017

https://arxiv.org/pdf/1608.07249v7.pdf

Why Tensorflow?

• Computational graph abstraction�
• Faster compile time�
• Tensorboard for visualization�
• Data and model parallelism�
• Embedding projector�
• Dynamic Computation Graphs�
• By Google :p

How it works?

• For efficient numerical computing in Python most libraries do expensive operation outside Python.�
• TensorFlow takes this a step further by defining a graph of interacting operations that run entirely outside Python.�
• This graph of interacting operations is known as Data flow graph.��

Data Flow Graph

Data Flow Graphs

• They describe mathematical computation with a directed graph of nodes & edges.�
• Nodes typically implement mathematical operations or data endpoints�
• Edges describe the input/output relationships between nodes�
• Data edges carry dynamically-sized multi-dimensional data arrays, or tensors�

Basic Terminologies and Functions

• tf.Variable:�A variable maintains state in the graph across calls to graph computation.�
• tf.placeholder:�Inserts a placeholder for a tensor that will be fed into the graph.�
• tf.device:�Context manager that specifies the default device to use for newly created op nodes�
• tf.name_scope:�Context manager that installs given name as a new name scope in the default graph�
• tf.app.flags.DEFINE_<data type>�Defines command line flags/arguments to take inputs�
• For more details please visit Tensor Flow API

Data Flow Graphs Example 1

Output:�[ 30.]

Data Flow Graphs Example 2

Output:�Addition: [ 30.]

Subtraction: [-10.]

Multiplication: [ 200.]

Division: [ 0.5]

​

Data Flow Graphs Example 3

Installation

CPU-only Version

• Simplest way using pip :��
• Other ways :
• Virtualenv
• Docker
• Anaconda
• From source

​

​

​

​

​

​

https://www.tensorflow.org/install/install_linux#determine_how_to_install_tensorflow

GPU Version

• Simplest way using pip :��
• Other ways :
• Virtualenv
• Docker
• Anaconda
• From source

​

​

​

​

​

​

https://www.tensorflow.org/install/install_linux#determine_how_to_install_tensorflow

NVIDIA requirements for GPU support

• CUDA Toolkit 8.0 : http://docs.nvidia.com/cuda/cuda-installation-guide-linux/#axzz4VZnqTJ2A �
• NVIDIA drivers�
• cuDNN v5.1 : https://developer.nvidia.com/cudnn �
• libcupti-dev library :

MNIST Demo

MNIST Demo Using Softmax Regression

• We will implement basic classification algorithm for handwriting recognition using MNIST (Modified National Institute Standards and Technology) dataset.

MNIST Dataset

• The downloaded dataset is split into 3 parts:
• Training data - 55,000 data points
• Test data - 10,000 data points
• Validation data - 5,000 data points

• Each data point consist of two parts: an image and a corresponding label�
• Image is an 28x28 matrix. We flatten it into an array of 784 length.

�

Basic Model for MNIST Problem

​

�

Data Flow Graph

Weights Visualization

MNIST Demo

Questions?

Important Links

• Softmax Regression code: https://github.com/aadilh/mnist-tensorflow �
• Tensorflow Installation: https://www.tensorflow.org/install/ �
• Tensorflow Tutorials: https://www.tensorflow.org/get_started/get_started �
• Benchmark Paper: https://arxiv.org/pdf/1608.07249v7.pdf �

​

Thanks!

Contact me:

Aadil Hayat�9616827469�hayataadil@gmail.com

​