AutoKeras Integration for Automated Neural Architecture Selection

AutoKeras integration for automatic neural network architecture selection

AutoKeras — Neural Architecture Search (NAS) on top of Keras/TensorFlow. Automatically searches for the optimal neural network architecture for a given task: number of layers, layer types, dimensions, dropout, and normalization.

Supported tasks

AutoKeras covers:

autokeras_tasks = {
    'ImageClassifier': 'классификация изображений — CNN архитектура',
    'ImageRegressor': 'регрессия на изображениях',
    'TextClassifier': 'классификация текста — Transformer/LSTM',
    'TextRegressor': 'регрессия на тексте',
    'StructuredDataClassifier': 'табличные данные — MLP + attention',
    'StructuredDataRegressor': 'регрессия на табличных данных',
    'TimeseriesForecaster': 'прогнозирование временных рядов',
    'MultiModal': 'комбинация типов данных'
}

AutoKeras for images

Search CNN architecture:

import autokeras as ak
import numpy as np
from sklearn.model_selection import train_test_split

def search_image_classifier(images: np.ndarray,
                              labels: np.ndarray,
                              max_trials: int = 30,
                              epochs: int = 20) -> dict:
    """
    images: (N, H, W, C) или (N, H, W)
    max_trials: количество архитектур для попытки
    """
    X_train, X_val, y_train, y_val = train_test_split(
        images, labels, test_size=0.2, random_state=42
    )

    clf = ak.ImageClassifier(
        overwrite=True,
        max_trials=max_trials,
        objective='val_accuracy',
        directory='/tmp/autokeras_image'
    )

    clf.fit(
        X_train, y_train,
        epochs=epochs,
        validation_data=(X_val, y_val),
        callbacks=[
            ak.callbacks.EarlyStopping(patience=5)
        ]
    )

    # Экспорт лучшей модели
    best_model = clf.export_model()
    val_accuracy = clf.evaluate(X_val, y_val)[1]

    return {
        'best_architecture': best_model.summary(),
        'val_accuracy': val_accuracy,
        'trials_evaluated': max_trials
    }

AutoKeras for text

Searching for Transformer/LSTM architecture:

def search_text_classifier(texts: list, labels: np.ndarray,
                             max_trials: int = 20) -> dict:
    """
    AutoKeras пробует: BERT fine-tuning, LSTM, Transformer с NAS.
    Автоматически токенизирует и создаёт embeddings.
    """
    clf = ak.TextClassifier(
        overwrite=True,
        max_trials=max_trials
    )

    clf.fit(texts, labels, epochs=10)
    best_model = clf.export_model()

    return {'model': best_model}

Custom Block Architecture

When standard tasks don't fit:

def custom_block_search():
    """
    AutoKeras Block API: задаём пространство поиска вручную.
    Полезно для нестандартных архитектур.
    """
    input_node = ak.ImageInput()

    # Блок: пробуем разные CNN backbone
    output_node = ak.ImageBlock(
        block_type='efficient',  # efficientnet или 'resnet', 'xception'
        augment=True
    )(input_node)

    # Блок: пробуем размеры Dense слоёв
    output_node = ak.DenseBlock(
        num_layers=ak.hyperparameters.Choice('dense_layers', [1, 2, 3]),
        use_batchnorm=True
    )(output_node)

    output_node = ak.ClassificationHead()(output_node)

    automodel = ak.AutoModel(
        inputs=input_node,
        outputs=output_node,
        max_trials=15,
        overwrite=True
    )
    return automodel

Export to TensorFlow Serving:

import tensorflow as tf

def export_for_serving(autokeras_model, export_path: str):
    """AutoKeras модель = Keras модель → стандартный экспорт"""
    tf.saved_model.save(autokeras_model, export_path)
    # TFLite для мобильных/edge
    converter = tf.lite.TFLiteConverter.from_saved_model(export_path)
    tflite_model = converter.convert()

    with open(f'{export_path}/model.tflite', 'wb') as f:
        f.write(tflite_model)

Deadlines: AutoKeras baseline for a standard task: 1-3 days. Custom Block API, export to TF Serving/TFLite, and multi-modal tasks: 1-2 weeks.