The Pattern

Three steps, every time

The Functional API follows a simple three-step pattern, and once it clicks it never changes no matter how complex the graph gets:

Define inputs with keras.Input(shape=...) — this creates a symbolic tensor, a placeholder that carries shape and dtype but no data yet.
Chain layers by calling them on tensors: x = Layer()(x). Each call records an edge in the graph.
Create model with keras.Model(inputs, outputs) — Keras traces backward from outputs to inputs and freezes the DAG.

The double-duty parentheses

The single idea that makes this whole API feel natural: every layer is a callable. The Code block below shows the parentheses doing double duty — layers.Dense(256, activation="relu") constructs a layer object, and the second (inputs) calls it on a tensor. Construction allocates the weights; calling wires the layer into the graph and returns a new symbolic tensor.

That separation is the payoff. Because constructing and calling are two distinct steps, you can hold a layer instance in a variable and call it more than once — and every call reuses the same weights. That single fact is what makes shared layers (next lessons) and weight tying possible at all. Notice too that nothing here runs a forward pass on real data: you're describing a graph, not executing it, which is exactly why the result is serializable.

Code

The three-step Functional pattern (MNIST classifier)·python

import keras
from keras import layers

# Step 1: Define input
inputs = keras.Input(shape=(784,))

# Step 2: Chain layers
x = layers.Dense(256, activation="relu")(inputs)
x = layers.Dropout(0.3)(x)
x = layers.Dense(128, activation="relu")(x)
outputs = layers.Dense(10, activation="softmax")(x)

# Step 3: Create model
model = keras.Model(inputs=inputs, outputs=outputs, name="my_classifier")
model.summary()

Three steps, every time

The double-duty parentheses

Code

External links

Exercise

Progress

Comments 0