AutoKeras — AutoML for Deep Learning with Keras

Introduction

AutoKeras is an AutoML library built on top of Keras that automates neural architecture search and hyperparameter tuning. It lets developers train high-quality deep learning models with minimal machine learning expertise by specifying only the task type and data.

What AutoKeras Does

• Searches for optimal neural network architectures using efficient NAS algorithms
• Automatically selects preprocessing, data augmentation, and normalization strategies
• Supports image classification, text classification, structured data, and regression tasks
• Tunes hyperparameters like learning rate, layer sizes, and dropout rates
• Exports the best model as a standard Keras model for deployment

Architecture Overview

AutoKeras uses Keras Tuner as its search backend. It defines a search space of possible architectures as a graph of interchangeable blocks (convolutional, dense, transformer). The tuner explores this space using Bayesian optimization or greedy search, training candidate models and evaluating them on a validation set. The best architecture is returned as a standard Keras Sequential or Functional model.

Self-Hosting & Configuration

• Install via pip; requires TensorFlow 2.x as the backend
• Set max_trials to control how many architectures are evaluated
• Use overwrite=True to restart a search or False to resume a previous one
• Configure epochs and validation_split in the fit() call
• Export the final model with export_model() for serving with TensorFlow Serving or ONNX

Key Features

• Task-oriented API covers image, text, structured data, and time-series problems
• Multi-modal input support combines images and tabular data in a single model
• Customizable search space lets advanced users constrain architecture blocks
• Integrated early stopping prevents wasted compute on poor candidates
• Exports standard Keras models compatible with the full TensorFlow ecosystem

Comparison with Similar Tools

• Auto-Sklearn — AutoML for scikit-learn classical models; AutoKeras targets deep learning architectures
• H2O AutoML — enterprise platform covering many model types; AutoKeras focuses on neural networks with Keras
• Ludwig — declarative config-based training; AutoKeras automates architecture search
• FLAML — fast lightweight AutoML; AutoKeras specializes in neural architecture search
• Google AutoML — managed cloud service; AutoKeras is open-source and runs locally

FAQ

Q: How long does an AutoKeras search take? A: Depends on max_trials and dataset size. A 10-trial search on MNIST finishes in minutes; larger datasets with more trials can take hours.

Q: Can I use a GPU? A: Yes. AutoKeras uses TensorFlow, which automatically detects and uses available GPUs.

Q: Is AutoKeras suitable for production models? A: The exported Keras model is a standard TensorFlow model that can be served with TF Serving, converted to TFLite, or exported to ONNX.

Q: Can I customize the search space? A: Yes. Use AutoModel with custom blocks from the autokeras.blocks module to define which layers and hyperparameters to search.

Sources

• https://github.com/keras-team/autokeras
• https://autokeras.com/