Auto-PyTorch

Auto-PyTorch is a free, self-hosted automated machine-learning toolkit based on PyTorch. It automates model selection and hyperparameter tuning, including neural-network architecture and training settings, within a user-set search budget. Its documented methods include Bayesian optimization, meta-learning, SMAC and Hyperband. Supported tasks include tabular classification, tabular regression and time-series forecasting. For tabular work, preprocessing can include imputation, categorical encoding, scaling and feature processing, with related settings tuned during the search. The toolkit can assemble ensembles from models’ validation-set predictions; users can configure ensemble size and candidate limits. Users can also set estimator memory limits and an overall wall-time limit. Parallel Bayesian optimization is available through Dask.distributed, though workers need shared filesystem access to training data and models. Installation options include PyPI, manual setup and a Docker image. The documented requirements are Linux, Python 3.7 or later, a C++11-capable compiler and SWIG 3.0; SWIG 4.0 or later is not supported. Forecasting needs additional dependencies. The project states that it uses the 3-clause BSD license.

Auto-PyTorch offers configurable automated search and ensembling for several machine-learning tasks. Check its Linux and SWIG requirements, and account for shared storage or extra forecasting dependencies where relevant.

Auto-PyTorch suits developers and machine-learning practitioners who want to automate model search in a self-hosted, code-based workflow. It is relevant to users working with tabular classification, regression or time-series forecasting.

Auto-PyTorch is a code-first AutoML toolkit for teams that want PyTorch-based model and pipeline search on their own Linux machines. It is best suited to Python practitioners who can manage dependencies and compute budgets. Its strongest case is configurable automation at no software cost; its weakest is a demanding, Linux-only installation with a strict SWIG version.

Overview

Auto-PyTorch searches machine-learning configurations instead of requiring users to select every algorithm and setting manually. Bayesian optimization, meta-learning, SMAC, and Hyperband explore pipelines within a user-defined budget, including neural-network architecture and training hyperparameters. The documented task coverage is tabular classification, tabular regression, and time-series forecasting.

This is self-hosted software with a code workflow, not a hosted visual service. That gives technical teams control over infrastructure and experiments, but it also puts environment setup, search configuration, and result evaluation on the user.

Key features

Automated pipeline and preprocessing search

For tabular data, preprocessing is searched alongside models. Imputation, categorical encoding, scaling, and other feature-preprocessing choices have tunable hyperparameters, reducing the amount of pipeline assembly done by hand. The benefit is greatest when teams want broad, repeatable exploration; users still need to understand the data and configure a meaningful evaluation budget.

Model search and ensembles

Auto-PyTorch builds ensembles by comparing model predictions on a validation set and selecting candidates accordingly. Ensemble size and candidate limits are configurable, so users can trade potential accuracy gains against runtime and complexity rather than accept a fixed recipe.

Budgets and parallelism

Separate controls let users cap estimator memory and total wall time for the search. Parallel Bayesian optimization uses Dask.distributed, which can shorten exploration on suitable infrastructure. Multiple workers must share access to training data and models through a shared file system, an operational requirement that matters for distributed deployments.

Forecasting and ecosystem

Time-series forecasting is supported, but it needs dependencies beyond the base installation. The documented integrations include PyTorch, scikit-learn transformers and column transformers, Dask.distributed, and threadpoolctl, making the toolkit most natural in an existing Python ML stack.

Pricing

Auto-PyTorch costs nothing under its free plan and uses the 3-clause BSD license. There is no paid tier or free trial to evaluate; the trade-off is that users receive a software toolkit rather than managed hosting, support tiers, or a graphical workflow. It fits individuals, researchers, and engineering teams able to supply their own machines and maintenance.

Platforms

Auto-PyTorch is Linux and self-hosted. Installation requires Python 3.7 or later, a C++11-capable compiler, and SWIG 3.0.*; SWIG 4.0 or later is not supported. A Docker image is documented, and installation is available through PyPI or a manual Python environment. These constraints make it a poor match for teams standardized on other operating systems or unwilling to pin an older SWIG release.

Who it's for

Choose Auto-PyTorch if you write Python, need automated selection and tuning for tabular or forecasting work, and want explicit control over memory, wall time, ensemble limits, and infrastructure. It is also a sensible fit for teams already using PyTorch and scikit-learn components. Look elsewhere if you need a browser-based workflow, broad desktop support, or a managed service that removes environment and distributed-storage administration.

Pros and cons

• Pros: Free BSD-licensed toolkit with no usage charge.
• Pros: Combines Bayesian optimization, meta-learning, SMAC, and Hyperband for configurable search.
• Pros: Tunes tabular preprocessing and supports configurable validation-based ensembles.
• Pros: User-set memory and wall-time limits help contain resource use.
• Cons: Linux-only support and a C++11 compiler requirement narrow deployment options.
• Cons: SWIG 4.0 or newer is unsupported, complicating modern environment management.
• Cons: Parallel workers require shared storage, and forecasting adds dependencies.
• Cons: Code-only, self-hosted operation demands more technical ownership than hosted AutoML.

Alternatives

FEDOT is another free, BSD-licensed AutoML framework with a free plan and support across API, Linux, macOS, Windows, and self-hosted setups. Pick it when broader platform coverage matters more than Auto-PyTorch's PyTorch-centered search.

AutoGluon is a free Apache 2.0 Python library for Linux, macOS, Windows, and self-hosted use; choose it when cross-platform open-source deployment is the priority.

AutoKeras is a free pip-installed Python package for common desktop and self-hosted platforms, suited to users who want that package's approach instead.

LightAutoML is a free Apache-licensed Python library installable from PyPI, with Linux, macOS, Windows, web, and self-hosted availability.

Amazon SageMaker Autopilot is a pay-as-you-go API and web service with a free trial; choose it when managed cloud operation is preferable to maintaining a local Linux stack.

BigML offers freemium API, web, macOS, and self-hosted access, including a free tier capped at 16 MB per task, two parallel tasks, and one user; choose it when those hosted and quota-based options fit better.

JADBio has a free Basic plan with one seat, three projects, 50 MB uploads, 500 MB storage, and one model export, plus paid Team pricing; pick it for a managed, project-oriented workflow.

Siemens Building X is a freemium platform spanning web, mobile, desktop, API, and Linux with yearly subscription billing; choose it for that broader application and API environment.

Verdict

Auto-PyTorch is a strong choice for technically capable Linux teams that want free, PyTorch-based automation with unusually direct control over search, ensembles, and resource limits. Choose it for configurable local experimentation; look elsewhere if strict SWIG and Linux requirements, shared-storage setup, or code-only operation are barriers.