Dissecting tf.function to Discover AutoGraph Strengths and Subtleties

Explore the intricacies of AutoGraph and tf.function in TensorFlow 2.0 through this EuroPython 2019 conference talk. Delve into the strengths and subtleties of these features, learning how to create and reuse graphs, handle state-creating functions, and optimize performance using tf.Tensor objects. Gain essential skills for writing efficient, graph-convertible code as the speaker guides you through common pitfalls and best practices. Discover the nuances of two-phase execution, dynamic vs. static typing, and the impact of tf.function on performance. By the end of this talk, develop a comprehensive understanding of AutoGraph's capabilities and limitations, enabling you to bridge the gap between TensorFlow 1.x and 2.0 effectively.

Syllabus

Introduction
Twophase execution
Problem description
Flow1 solution
Adding the decorator
First output
Exception
Lesson
Solution
TFfunction are not graph convertible
Dynamically typed vs statically typed
Identity
First test
Analysis
Code
Weird behavior
Summary
F function outers
Performance measurement
Use tf tensor everywhere
What happens when we plug into tffunction
Problems with tffunction
Recap
Questions