Functional architectures
Build a directed model graph from one or more input tensors and computed output tensors.
TensorFlow · Class
keras.Model
tf.keras.Model groups layers into a callable object with training and inference features. Create it with the Functional API or by subclassing it.
How it works
tf.keras.Model groups layers into a callable model that supports training and inference. You can construct it from input/output tensors with the Functional API or subclass it by defining layers in __init__() and the forward pass in call().
Syntax
tf.keras.Model(*args, **kwargs)Parameters
| Name | Type | Description |
|---|---|---|
*args, **kwargs | Any | Positional and keyword arguments accepted by the Model constructor. |
Returns
A Keras Model object grouping layers and exposing training and inference features.
Key points
- Model combines layers with built-in capabilities for training and inference.
- The documentation describes three construction approaches, including the Functional API and subclassing.
- The Functional API expresses a forward pass by chaining layers from Input tensors.
- Finish a functional model by passing its input and output tensors to the constructor.
- Use compile() to configure the losses and metrics used by the model.
- Use fit() for training and predict() for generating predictions.
Common mistakes
- When subclassing Model, put layer definitions in __init__() and the forward computation in call().
- Configure losses, metrics, and the optimizer with compile() before training.
Example: Functional model
import tensorflow as tf
from tensorflow import keras
inputs = keras.Input(shape=(37,))
x = keras.layers.Dense(32, activation="relu")(inputs)
outputs = keras.layers.Dense(5, activation="softmax")(x)
model = keras.Model(inputs=inputs, outputs=outputs)
model.compile(optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"])
model.summary()Creates a functional model, configures its training behavior, and prints its architecture.
When to use it
Subclassed models
Implement custom forward-pass behavior by defining layers and call().
End-to-end model workflows
Configure, train, evaluate, and use neural networks for prediction.
Interview questions
QWhat is tf.keras.Model?
A Model groups layers and provides training and inference functionality. It can be built through the Functional API or by subclassing Layer-like behavior.
Follow-up: How would you choose between Functional API and subclassing? Use the Functional API for graph-shaped models with explicit inputs and outputs; subclass when custom forward-pass logic is needed.
QHow do you create a functional Model?
The Functional API starts with Input tensors, applies layers to create outputs, and constructs the model from those inputs and outputs. The resulting model remains callable like a layer.
Follow-up: Can models be nested? Yes. A model can be used as a layer, so one model can contain another model.
QHow is a Model trained and used for inference?
After construction, call compile() to configure losses and metrics, fit() to train, and predict() to generate predictions. These are built-in model operations.
Follow-up: What other standard operation evaluates a trained model? evaluate() returns loss and metric values configured through compile().
Related
tf.keras.layers.Layertf.keras.Inputtf.keras.Sequentialtf.keras.Model.compiletf.keras.Model.fittf.keras.Model.predicttf.keras.Model.evaluate