KerasTuner automates the repetitive search for a high-performing Keras model configuration. Because hyperparameter choices are generally nondifferentiable, each candidate requires a separate training run, and small metric differences can be noisy. section 18.1.1
A small MNIST search
Replace fixed values such as units=32 with searchable choices such as hp.Int("units", 16, 64, step=16), and let hp.Choice select an optimizer. A model-building function receives hp, constructs and compiles the model, and returns it. A BayesianOptimization tuner can then search for the configuration that maximizes val_accuracy. section 18.1.1
Listing 18.1 A KerasTuner model-building function
import keras from keras import layers def build_model(hp): units = hp.Int(name="units", min_value=16, max_value=64, step=16) #1 model = keras.Sequential( [ layers.Dense(units, activation="relu"), layers.Dense(10, activation="softmax"), ] ) optimizer = hp.Choice(name="optimizer", values=["rmsprop", "adam"])#2 model.compile( optimizer=optimizer, loss="sparse_categorical_crossentropy", metrics=["accuracy"], ) return model #3
Inspect the concrete model-building function that declares integer and categorical hyperparameters before compiling the Keras model. section 18.1.1
A practical KerasTuner workflow is:
- Define the search space with hp.Int, hp.Choice, or similar hyperparameter ranges. section 18.1.1
- Pass the model builder to a tuner such as RandomSearch, BayesianOptimization, or Hyperband. section 18.1.1
- Set an objective, limit trials, and optionally repeat executions per trial to account for noisy training results. section 18.1.1
- Call tuner.search with training and validation data, then retrieve the best hyperparameters or trained models. section 18.1.1
tuner = kt.BayesianOptimization( build_model, #1 objective="val_accuracy", #2 max_trials=20, #3 executions_per_trial=2, #4 directory="mnist_kt_test", #5 overwrite=True, #6 )
Inspect a complete BayesianOptimization configuration, including the validation objective, trial budget, repeated executions, and output directory. section 18.1.1
Treat the validation score as a tuning signal rather than a final guarantee: repeated optimization against one validation set can overfit that set. Retrain promising configurations with a suitable epoch count and evaluate on held-out test data. section 18.4 section 18.1.1
