Past Event: CSEM Student Forum

Tensorflow: A Short Primer

Sheroze Sheriffdeen, ICES, UT Austin

Abstract

TensorFlow is an open-source machine learning library for research and production. This primer will introduce TensorFlow’s high-level interfaces which provide building blocks to create and train deep learning models, pre-made and custom Estimators to write your own models, and low-level interfaces to manipulate the underlying computational graphs. This primer will introduce Keras, a high-level API to build and train deep learning models. It is used for fast prototyping, advanced research, and production. It is user friendly, modular, and easy to extend. Keras provides a simple interface optimized for common use cases and its models are made by connecting configurable building building blocks together. TensorFlow’s Estimators are high-level representations of complete models. It handles the details of initialization, logging, saving and restoring, and many other features so you can concentrate on your model. This primer will show examples of pre-made Estimators, custom Estimators, and how to define your input functions to feed data into these models. Finally, this primer will cover TensorFlow’s low-level interface which focuses on building a dataflow graph. In a dataflow graph, the nodes represent units of computation, and the edges represent the data consumed or produced by a computation. This leads to a low-level programming model in which you first define the graph, then create a TensorFlow session to run parts of the graph across a set of local and remote devices. It is important to understand this model as it aids in debugging and extending high-level deep learning models.