As we saw in the previous chapter, TensorFlow provides powerful tools for tensor manipulation and automatic differentiation using tf.GradientTape. While you could build complex models directly using these fundamental operations, tracking variables, gradients, and computation graphs manually can become intricate and error-prone, especially for deep networks.

This is where Keras comes in. Keras is a high-level API for building and training neural networks, and it's deeply integrated within TensorFlow 2.x. Think of Keras as a user-friendly interface that sits on top of TensorFlow's core functionalities. It allows you to define, train, and evaluate models with significantly less code and cognitive overhead compared to using lower-level TensorFlow operations directly.

Why Use Keras?

The primary goal of Keras is to make deep learning development faster and easier. It achieves this through several design principles:

User Friendliness: Keras APIs are designed to be intuitive and consistent. Common neural network components like layers, activation functions, optimizers, and loss functions are readily available as pre-built, configurable modules. This reduces the amount of boilerplate code you need to write.
Modularity: Keras models are typically constructed by connecting configurable building blocks (like layers). These components can be easily combined and reused, promoting organized model development.
Extensibility: While Keras provides many standard components, it also makes it straightforward to create custom layers, loss functions, and metrics for research or specialized applications. We'll touch upon this later in the chapter.
Tight TensorFlow Integration: Since Keras is TensorFlow's official high-level API, it works naturally with other TensorFlow features like tf.data for efficient data pipelines and tf.function for graph optimization. You can easily mix Keras components with lower-level TensorFlow code when needed.

Keras: Not Just a Wrapper

It's important to understand that Keras isn't merely a wrapper around TensorFlow; it is the standard way to build models in TensorFlow for most users. When you use tensorflow.keras, you are using TensorFlow. Keras provides the abstractions (like Layer, Model, Sequential) that translate your model definitions into the underlying TensorFlow computation graph and operations.

Consider the relationship like this:

Keras provides a high-level abstraction layer over TensorFlow's core operations, simplifying model development.

By using Keras, you leverage TensorFlow's performance optimizations (like graph execution via tf.function) and hardware acceleration capabilities (CPU, GPU, TPU) without needing to manage those details directly in most cases.

In the following sections, we will explore the primary ways Keras allows you to define models: the Sequential API for simple linear stacks of layers, and the Functional API for building more complex architectures with multiple inputs, outputs, or shared layers.