This tutorial will focus on practical aspects of implementing and training deep learning models. We will follow, step by step, the implementation of linear and nonlinear models with TensorFlow, focusing on the impact of tuning hyperparameters and architectures. To keep track of results we will use the visualization tool TensorBoard and show its functionalities.
The goal of this tutorial is for you to learn how to quickly build models and processing pipelines in TensorFlow. You will apply them to a standard dataset, and keep track of the experiments' performance with TensorBoard, a visualization tool. We will explain in detail how to implement the different parameters and architectures in TensorFlow in a simple way.
We strongly recommend to have covered some machine learning basic concepts previously, as we wont cover them during the tutorial:
The tutorial is based on Jupyter notebooks that will be made public before the tutorial. They contain code and explanations that the attendees can download, execute and modify during the talk. It will comprise the following parts:
(10 min) We will provide a notebook containing the instructions to install the libraries used during the tutorial and to check they work correctly. We recommend this to be done in advance before the talk.
(20 min) We will cover the implementation of a linear model to exemplify the different components of the TensorFlow pipeline. This includes compilation of models and how TensorFlow is executing your code.
(25 min) We will extend the linear model previously built into a deep neural network by using the Layers module. During this section, we will also describe the possible problems of training big models and how to overcome them.
(20 min) We will present the TensorBoard interface and how we can use it to make the experimentation process easier by collecting and visualizing the results.
(15 minutes) We allocate this time for attendees to play with the notebooks creating new models and trying to optimize the performance on the given dataset. Both instructors will be answering questions.