Hyperparameter tuning selects configuration values by training and evaluating models across alternatives. With TensorFlow and Keras Tuner, a model-building function exposes tunable parameters such as layer width and learning rate. section 8.1.1
A practical tuning workflow:
- Define search ranges, such as 32–512 dense-layer units and Adam learning rates of 0.01, 0.001, or 0.0001. section 8.1.1
- Select a strategy such as Hyperband, which combines adaptive resource allocation with early stopping. section 8.1.1
- Search against validation accuracy, stop unpromising training with validation-loss patience, then rebuild and train the best model. section 8.1.1
MNIST with Keras Tuner
A TensorFlow/Keras model tunes its first dense layer and Adam learning rate over MNIST, searches with Hyperband, and evaluates the resulting model on test data. section 8.1.1
Listing 8.16 The Hyperband model tuner
tuner = kt.Hyperband(model_builder,
objective='val_accuracy',
max_epochs=10,
factor=3,
directory='my_dir',
project_name='intro_to_kt')
Inspect the Hyperband tuner configuration, including the validation objective, epoch limit, search factor, and output directory. section 8.1.1
