How aeon-toolkit/aeon works
Project Type and Purpose
• aeon is an open-source Python toolkit for time series machine learning, providing algorithms for classification, regression, clustering, forecasting, anomaly detection, transformations, similarity search, visualization, benchmarking, and more.
• It follows a plugin-style architecture: core base classes define the API and contract, with concrete estimators and transformers implemented in subpackages.
High-Level Component Breakdown
a. Core Base Layer (highlighting base classes)
• aeon.base._base (BaseEstimator, BaseSeries, BaseCollection)
• aeon.base._compose (pipeline/ensemble composition logic)
• aeon.pipeline._make_pipeline, aeon.pipeline._sklearn_to_aeon
b. Algorithm Modules
• classification (distance-based, dictionary-based, deep_learning, feature-based, hybrid, interval-based, shapelet-based, sklearn wrappers, etc.)
• regression (same structure as classification)
• clustering
• forecasting
• anomaly_detection
• segmentation
• similarity_search
• transformations (series & collection)
• benchmarking (metrics, resampling, published_results loaders)
c. Data & IO Layer
• aeon.datasets: data loaders, single-problem loaders, time series anomaly detection datasets
• file formats (.ts, .arff, .csv)
d. Utility & Support Packages
• aeon.utils (conversion, discovery, numba helpers, validation, tags, network weight norm, show_versions)
• aeon.testing (data generation, estimator checking, mock_estimators, expected_results)
e. Deep-Learning Network Layer
• aeon.networks: implementations of CNN, RNN, autoencoders, etc., used by classification/regression/deep modules
f. Visualization Layer
• aeon.visualisation: estimator plotting, learning task plotting, results plotting, series/collection plotting
g. External Integrations
• scikit-learn compatibility
• numba JIT for performance
• binder, GitHub Actions, ReadTheDocs, PyPI/Conda packaging
Relationships & Data Flows
• Pipeline Layer orchestrates Algorithm Modules (composition, sklearn interop).
• Testing Layer targets every module (estimator_checking, unit tests).
• Documentation Generation (docs folder, Sphinx) reads code and README.
• CI/CD (GitHub Actions) triggers tests, type checks, codecov, release workflows.
• Packaging & Distribution (pyproject.toml → PyPI/Conda).
Architectural Patterns & Design Principles
• Inheritance & Polymorphism: all estimators extend core base classes to adhere to a uniform API.
• Composition: pipelines and ensemble constructors reside in _compose modules.
• Plugin/Registration: base._register and utils.discovery enable automatic estimator discovery.
• Layered Architecture: Data→Core API→Algorithm Implementations→Visualization/Benchmarking→Presentation.
• Separation of Concerns: each subpackage has its own tests directory; testing and CI are decoupled.
• Components:
– Data & IO (cyan)
– Utilities & Testing (orange)
– Network Layer (violet)
– Visualization & Benchmarking (lime)
– CI/CD & Packaging (purple)
– External Services (grey)
• Labels:
– Clearly label package names and key classes (e.g., “BaseEstimator”).
Highlighting Base Classes Usage
• Subclass hierarchy fans: classification, regression, clustering, forecasting, anomaly_detection, segmentation, similarity_search, transformations all inherit from BaseSeries or BaseCollection.
• “_register” and “discovery” utilities linking new subclass registration to the central registry.