KerasTuner HyperModels

The HyperModel base class makes the search space better encapsulated for sharing and reuse. A HyperModel subclass only needs to implement a build(self, hp) method, which creates a keras.Model using the hp argument to define the hyperparameters and returns the model instance. A simple code example is shown as follows.

class MyHyperModel(kt.HyperModel):
  def build(self, hp):
    model = keras.Sequential()
    model.add(keras.layers.Dense(
        hp.Choice('units', [8, 16, 32]),
        activation='relu'))
    model.add(keras.layers.Dense(1, activation='relu'))
    model.compile(loss='mse')
    return model

You can pass a HyperModel instance to the Tuner as the search space.

tuner = kt.RandomSearch(
    MyHyperModel(),
    objective='val_loss',
    max_trials=5)

There are also some built-in HyperModel subclasses (e.g. HyperResNet, HyperXception) for the users to directly use so that the users don't need to write their own search spaces.

tuner = kt.RandomSearch(
    HyperResNet(input_shape=(28, 28, 1), classes=10),
    objective='val_loss',
    max_trials=5)

{{toc}}