Skip to the content
Know your tech before you buy it. Makers: get listed from $49/mo

AutoKeras

autokeras.com

  • Free plan
  • Open source
  • Windows
  • macOS
  • Linux
  • API
  • Self-hosted
8.1 Club score Visit autokeras.com

What AutoKeras is

AutoKeras is a free, open-source AutoML system based on Keras that lets users build machine-learning models through code. It supports image and text classification and regression, as well as classification and regression for structured data. A documentation tutorial covers multi-task and multimodal datasets. Users can assemble custom models from AutoKeras blocks and let the system search for detailed configurations. For image work, architecture type, normalization, and data augmentation can be set manually or left for automatic tuning. AutoKeras can export the best model it finds 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 2.3.0 or later with the PyTorch backend. By default, the final 20% of training data is reserved for validation, though users can select another split or provide validation data. Model selection uses validation loss by default; metrics and objectives can be customized.

Who it suits

It suits developers and machine-learning practitioners who want code-based automated model selection and are comfortable installing and configuring a Python package.

Good to know, and what to watch

  • Free and open source.
  • Supports image, text, and structured-data tasks.
  • Can export the selected model as a Keras Model.
  • Users can customize metrics and objectives.
  • Workflow is code-based.
  • Installation requires Python package setup.
  • Validation defaults to the final 20% of training data.

AutoKeras

Free

Open-source library

Free

Open source

Getting started

  1. Open the AutoKeras website
  2. Install the Python package with pip
  3. Use Python 3 in your environment
  4. Choose a supported task and build a model from AutoKeras blocks
  5. Run training with the default or a custom validation split
  6. Export the selected result as a Keras Model

The Club's review

Overview

AutoKeras is a free, open-source Python package for automating machine-learning model design and selection. It suits developers, researchers, and data scientists who are comfortable working in code and want one toolkit for image, text, and structured-data problems. The trade-off is clear: its flexibility comes through Python configuration rather than a graphical workflow.

Key features

AutoKeras covers image and text classification and regression, plus classification and regression for structured data. That breadth is practical for teams handling several supervised-learning workloads because the same package can support different data types. Documentation also includes multi-task and multimodal examples, useful when a project combines outputs or modalities, although those scenarios remain code-led.

Model construction is configurable rather than all-or-nothing. You can assemble a model from AutoKeras blocks while allowing the search process to tune detailed configurations. For image work, architecture type, normalization, and data augmentation can be chosen explicitly or handed to automated tuning. This gives experienced users meaningful control without forcing them to specify every configuration, but beginners still need to understand the underlying modeling choices.

AutoKeras exports the best result as a Keras Model, making the search output usable in a standard Keras workflow. Validation behavior is also adjustable: the default reserves the last 20% of training data for validation, while users can supply another split or their own validation data. Model selection defaults to validation loss, with support for custom metrics and objectives, so the chosen model can reflect a project's actual success criterion rather than a fixed default.

Repeated runs can be resumed. With overwrite set to false, running the same code again continues a previous fit from the contents of its directory. That is useful for interrupted or iterative searches, provided the run directory is maintained consistently.

Pricing

AutoKeras costs nothing. The AutoKeras plan is the open-source Python package installed with pip, with no software charge, seat fee, subscription term, or usage quota stated. This makes it suitable for individual experiments and teams that can host and operate their own environment. The price advantage does not remove the need to provide your own compute, storage, maintenance, and technical support.

Platforms

AutoKeras runs on Linux, macOS, and Windows, and is self-hosted rather than delivered as a hosted application. Installation is provided through pip for the Python package. The project supports Python 3 and recommends PyTorch 2.3.0 or later when using the PyTorch backend. Questions are directed to GitHub Discussions, so support is community-oriented rather than a stated commercial service level.

Who it's for

Choose AutoKeras if you write Python, need automated feature engineering and model selection, and want to retain control over model blocks, validation, and optimization objectives. It is particularly appropriate for practitioners moving between image, text, and tabular projects, or for teams that need an exported Keras artifact.

It is a poor fit for users seeking a browser-based workflow, managed hosting, or a point-and-click experience. Organizations requiring stated security certifications or compliance commitments should also evaluate alternatives, since no such claims are provided for AutoKeras.

Pros and cons

Pros

  • No license cost: the package is free and open source, removing subscription and per-seat software charges.
  • Broad task coverage: image, text, and structured-data classification and regression are supported in one code-based toolkit.
  • Controlled automation: custom blocks, image settings, validation data, metrics, and objectives let advanced users shape the search.
  • Portable result: the selected model can be exported as a Keras Model.
  • Resumable searches: rerunning with overwrite disabled can continue an earlier fit from its directory.

Cons

  • Code-only workflow: there is no stated visual interface, which raises the skill requirement for first-time users.
  • Self-hosted operations: users must manage the runtime and computing environment themselves.
  • Community support: questions go through GitHub Discussions rather than a stated paid support SLA.
  • No stated compliance claims: regulated teams cannot rely on published security certifications from the project.

Alternatives

AutoML Software is the broader category overview for comparing automated machine-learning tools.

Pick BigML instead if you want a web, API, or self-hosted service with a free tier, 16 MB per-task datasets, two parallel tasks, and one user.

JADBio is an alternative when a freemium, web or API offering and defined project, storage, upload, and export limits matter.

Choose Amazon SageMaker Autopilot for a paid, API or web service billed pay-as-you-go rather than a local Python package.

FEDOT, LightAutoML, MLJAR AutoML, Auto-PyTorch, and AutoGluon are other open-source AutoML options.

Verdict

AutoKeras is a strong choice for Python practitioners who want free, self-hosted automated search across several data types while keeping control over model construction and selection criteria. Look elsewhere if your priority is a managed visual workflow, contractual support, or published compliance assurances.

The verdict

AutoKeras offers automated model search across several data types while allowing custom model blocks and selection objectives. It is a code-oriented option rather than a workflow interface.

Questions people ask

Is AutoKeras free?

Yes. The AutoKeras Python package is listed as free and is open source.

Which operating systems are supported?

AutoKeras supports Linux, macOS, Windows, and self-hosted use.

What tasks can it handle?

It supports image and text classification and regression, plus structured-data classification and regression.

How does it choose a model?

Validation loss is the default selection criterion. Custom metrics and objectives can be used instead.

Can a training run resume?

Yes. With overwrite set to false, rerunning the same code continues from the existing directory contents.

What the maker publishes

Community funding
Open Collective reports $2,065.59 raised and $205.67 disbursed for AutoKeras. source
Community size
Open Collective lists 32 contributors. source
Company location
The Texas A&M Engineering page identifies the institution in College Station, Texas. source
Custom search spaces
Users can connect built-in blocks or implement their own Block subclasses to customize the search space. source
Deployment ecosystem
The paper states exported models can be deployed through TensorFlow ecosystem tooling. source
Documentation depth
The documentation includes task tutorials, API reference pages, customization guidance, export instructions and an FAQ. source
Funding
Texas A&M reported that the project had been funded by DARPA through an AFRL contract. source
Image classification
The task API supports image-classification workflows. source
Image features
Image workflows can configure architecture type, normalization, and data augmentation, or leave choices to be tuned automatically. source
Image regression
The task API supports image-regression workflows. source
Installation
The project provides pip installation instructions for the Python package. source
Installation
The official installation method is pip install autokeras, with Python 3 required. source
Latest package release
PyPI lists version 3.0.0 as released on November 12, 2025. source
Licence
The source repository and package metadata specify Apache License 2.0. source
Limit: structured-data production fit
The same Reddit discussion reported relying on boosted trees for tabular data and not finding neural-network success in production. source
Maintainer
The package is developed by DATA Lab at Texas A&M University in collaboration with keras-team for version 1.0 and above. source
Maker
The package is developed by DATA Lab at Texas A&M University with Keras-team collaboration. source
Model design
Users can build custom models from AutoKeras blocks while letting the system search for detailed configurations. source
Model export
AutoKeras can export the best model found as a Keras Model. source
Model export
The best searched model can be exported as a Keras model. source
Model selection
AutoKeras uses validation loss by default to select the best model, and supports customized metrics and objectives. source
Multimodal
The documentation includes a tutorial for working with multi-task or multi-modal datasets. source
Multi-modal learning
AutoModel can combine different input types such as images and text. source
Multi-task learning
AutoModel can produce multiple prediction heads, including classification and regression outputs. source

Checked against the maker's pages on 15 Oct 2026.