Auto-PyTorch vs AutoKeras

Auto-PyTorch is a free, self-hosted machine-learning toolkit based on PyTorch. It automates algorithm selection and hyperparameter tuning, including neural-network architecture and training settings, within a user-defined search budget. Documented search methods include Bayesian optimisation, meta-learning, SMAC and Hyperband. Supported work includes tabular classification, tabular regression and time-series forecasting. Tabular preprocessing can perform imputation, categorical encoding, scaling and feature processing, while related hyperparameters are tuned. The toolkit can build ensembles from validation-set predictions, with controls for ensemble size and candidate limits. Users can set estimator memory limits and total wall time. Parallel Bayesian optimisation uses Dask.distributed; workers need shared filesystem access to training data and models. Installation documentation specifies Linux, Python 3.7 or later, a C++11-capable compiler and SWIG 3.0. SWIG 4.0 or later is unsupported. Forecasting needs extra dependencies. A Docker image and PyPI installation are documented, under the 3-clause BSD license.

AutoKeras is a free, open-source AutoML system based on Keras, designed to make machine learning accessible through code. It supports image and text classification and regression, as well as classification and regression for structured data. Its documentation also includes a tutorial for multi-task or multimodal datasets. Users can assemble custom models from AutoKeras blocks and let the system search for detailed configurations. Image workflows can set architecture type, normalization and data augmentation, or leave those choices for automatic tuning. The best model found can be exported as a Keras Model. The Python package is installed with pip and supports Linux, macOS and Windows, including self-hosted use. The installation guide supports Python 3 and recommends PyTorch version 2.3.0 or later with the PyTorch backend. By default, the final 20% of training data is used for validation, though users can set a different split or supply validation data. Model selection uses validation loss by default, with support for customized metrics and objectives. The project directs users to GitHub Discussions for questions.