Tensorflow is a numeric computation framework open-sourced by Google. From a system’s perspective, Tensorflow’s distinguishing feature is the ability to parallelize and distribute computations over thousands of CPUs and GPUs, making it suitable for deployments on large clusters. From a developer’s perspective TensorFlow makes life easy: install TF on your laptop in secs, write and test your code there, then copy to server. The same code runs on your laptop or a massive cluster of thousands of nodes without modifications. Moreover, TF auto-differentiates all the functions it understands (infinite through composition) saving you hundreds of code lines and hours of debugging. All these are made possible by the simple principle of separating computation declaration from execution.

The course is an introduction to TF and computation graphs using TF’s Python API. Although introductory, the course covers a substantial breadth of TF’s concepts and tools with hands-on examples and exercises.

Objectives

1. Learn the basics of declarative computation (computation graphs).
2. Learn how TF implements declarative computation.
3. Learn how to implement basic neural networks for classification and regression tasks.

Who Should Attend

Data engineers that need a tool for building distributed workflows of complex data transformations.

Data scientists that want to use neural networks for regression and classification tasks.

Technical managers involved in the evaluation of technologies and human resource skills related to analytics and big data.

Prerequisites

1. Basic knowledge of Python and the numpy package
2. Basic understanding of what a regression and classification task is and what is training with examples.
3. Have a laptop with Ubuntu 16.04 or windows 10 OS, at least 4GB RAM and 32GB disk storage

Course Outline

1. Tensor Basics
   The computational graph model – imperative vs declarative programming
   Basic operations with tensors
   Graph Inspection & Visualization with TensorBoard
2. Linear Algebra with TF
   Operations on sparse and dense matrices and vectors
   Kronecker Products in TF
   From matrices to tensors
   Tensor tiling: The map Operator
   Reductions on tensors
   Auto-differentiation: TF’s distinguishing feature
3. Tensorflow I/O
   Feeding data with placeholders
   Exporting results
   Splitting and processing inputs with minibatches
4. Numerical optimization In TensorFlow
   Creating a symbolic objective function
   Gradient computations: building a simple gradient-descent optimizer
   Tweaking gradients: the compute_gradient and apply_gradient methods
   Predefined TF optimizers: gradient descent, adagrad and adam optimizers
   Predefined loss functions
   Building a linear regression model with TF in 3 lines
5. Introduction to Neural Networks and deep learning
   Fundamentals of Neural Nets
   Stochastic gradient descent
   The backpropagation algorithm
   Building a neural network classifier
   Learning embeddings of categorical variables