Inside TensorFlow - Functions, Not Sessions

Dive into an in-depth technical session from the TensorFlow team exploring the evolution from tf.Session to tf.function. Learn about the history, drawbacks, and motivations behind this transition, as well as the key classes and functions used to generate graphs from tf.functions. Explore topics such as Sessions, Runtime, Partial Run, TensorFlow Functions, Variables, and their compromises. Gain insights into the solution and its potential downsides. Follow a walkthrough covering Photograph Capture, Phonegraph, Automatic Control, Calling Functions, Differentiating Functions, and Concrete Functions. Examine the code base, Variable Lifting, Variable Subclasses, Variable Capture Scopes, Stateless Functions, and Lift to Graph concepts. Enhance your understanding of TensorFlow's internal workings and development process through this comprehensive 46-minute presentation by TensorFlow Software Engineer Alexandre Passos.

Introduction
Session vs Runtime
Session Not Run
Partial Run
TensorFlow Functions
Variables
The compromise
The solution
The downsides
Walkthrough
Photograph
Capture
Phonegraph
Automatic Control
Calling Functions
Differentiating Functions
Concrete Functions
Code Base
Variable Lifting
Variable Subclasses
Variable Capture Scopes
Stateless Functions
Lift to Graph