Keras Tuner Legacy Imports

1.0.5 · active · verified Thu Apr 16

The `kt-legacy` library provides backward-compatible import names for Keras Tuner. It allows users to import Keras Tuner components using the `kerastuner` namespace, which was the original import path, instead of the current `keras_tuner` namespace. This is particularly useful for migrating or maintaining compatibility with older codebases that expect the `kerastuner` module.

Common errors

```
ModuleNotFoundError: No module named 'kerastuner'
```
cause This error occurs when old code attempts to `import kerastuner` but the `kt-legacy` package is not installed, or `keras-tuner` itself is not installed. Modern `keras-tuner` installations use `keras_tuner` as the primary import path.

fix Install `kt-legacy` using `pip install kt-legacy`. Also ensure `keras-tuner` is installed via `pip install keras-tuner`.
```
AttributeError: module 'kerastuner' has no attribute 'RandomSearch'
```
cause This usually indicates that `kt-legacy` might be installed, but the underlying `keras-tuner` library is either missing or an incompatible version, preventing the `kerastuner` alias from correctly resolving to the actual Keras Tuner classes.

fix Verify that `keras-tuner` is installed and up-to-date (`pip install --upgrade keras-tuner`). If the problem persists, try reinstalling both: `pip install --upgrade keras-tuner kt-legacy`.

Warnings

deprecated The `kerastuner` import path is deprecated in favor of `keras_tuner` within the main Keras Tuner library. While `kt-legacy` provides compatibility, new projects should use `keras_tuner` directly to align with current best practices and avoid potential future breakage.
Fix: For new code, replace `import kerastuner` with `import keras_tuner` and update references accordingly. For existing code, ensure `kt-legacy` is installed alongside `keras-tuner`.
gotcha `kt-legacy` only provides the legacy import alias. It does not include the Keras Tuner functionality itself. `keras-tuner` must be installed separately for the aliased imports to function.
Fix: Ensure both `keras-tuner` and `kt-legacy` are installed: `pip install keras-tuner kt-legacy`.

Install

pip install kt-legacy Install `kt-legacy`
pip install keras-tuner kt-legacy Install `keras-tuner` and `kt-legacy` (recommended)

Imports

kerastuner
wrong:
```
import kerastuner as kt
```
correct:
```
import kerastuner as kt
```
The `kt-legacy` package enables `import kerastuner` for projects written with older Keras Tuner import paths. For new development, it's recommended to `import keras_tuner as kt` directly.

Quickstart

This quickstart demonstrates how to import and initialize a Keras Tuner `RandomSearch` tuner using the `kerastuner` legacy import provided by the `kt-legacy` package. It defines a simple hypermodel and initializes the tuner, ready for a hyperparameter search.

import keras
import kerastuner as kt

def build_model(hp):
    model = keras.Sequential([
        keras.layers.Flatten(input_shape=(28, 28)),
        keras.layers.Dense(
            units=hp.Int('units', min_value=32, max_value=512, step=32),
            activation='relu'
        ),
        keras.layers.Dense(10, activation='softmax')
    ])
    model.compile(
        optimizer=keras.optimizers.Adam(learning_rate=hp.Choice('learning_rate', values=[1e-2, 1e-3, 1e-4])),
        loss='sparse_categorical_crossentropy',
        metrics=['accuracy']
    )
    return model

# Example of using a tuner with the legacy import
tuner = kt.RandomSearch(
    hypermodel=build_model,
    objective='val_accuracy',
    max_trials=2, # For quick demonstration
    executions_per_trial=1,
    directory='my_dir', project_name='intro_to_kt_legacy'
)

# Note: To run search, you would typically need training data, e.g.,
# (img_train, label_train), (img_test, label_test) = keras.datasets.fashion_mnist.load_data()
# img_train = img_train.astype('float32') / 255.0
# label_train = label_train[:100] # Subset for quick example
# img_train = img_train[:100]
# tuner.search(img_train, label_train, epochs=2, validation_split=0.2)

print("Keras Tuner (legacy import) initialized successfully.")
# print(tuner.get_best_hyperparameters()[0].values)

view raw JSON →