Inside TensorFlow - tf.Keras - Part 1

Explore an in-depth technical overview of tf.Keras in this 54-minute video presentation by Francois Chollet, the creator of Keras. Dive into the internal architecture of Keras, understanding the role and functionality of layers, models, and the Functional API. Learn about lazy layer building, nested layers, loss collection, serialization, and the differences between eager and graph execution. Discover advanced topics such as automatic masking, symbolic inputs, and dynamic layers. Gain insider knowledge from the TensorFlow team's internal training session, offering valuable insights for both beginners and experienced developers working with TensorFlow and Keras.

Intro
The Keras architecture
What does a Layer do?
What does a Layer not do?
The most basic layer
A canonical lazy layer (build(), add_weight())
Nested layers
Basic usage of a layer
Defining losses on the fly and collecting them at the end
Making your layers serializable
Special call argument: training
Basic Model
A Model handles top-level functionality
Eager & graph execution for fit(), evaluate()
The Functional API is a way to create DAGs of layers
A Functional Model behaves like any other Layer/Model, but it has several methods autogenerated (call, build, get_config)
Anatomy of a Functional Model
keras history is the coordinates of the tensor in a 3D construction grid
Static input compatibility checks
Whole-model saving / serialization and reinstantiation across platforms
Automatic masking: a first example
Automatic masking: details
In-depth: what happens when you call a layer on symbolic inputs
Using dynamic layers