Apeiron¶
A PyTorch framework for continual learning that detects concept drift in a live data stream and adapts the model in place.
Apeiron runs a monitoring loop over a changing data stream, watches an evaluation metric, hands that metric to a drift detector, and — when the detector fires — pauses monitoring to run a continual-learning update on the model before resuming.
uv run python -m src.main --config examples/mnist/mnist.toml
The loop¶
Evaluate the current model on stream batches.
Aggregate the monitored metric at a configured interval.
Run a drift detector on the aggregated metric.
On drift, pause monitoring and run a continual-learning update loop.
Resume monitoring on the updated model until stream limits are reached.
Install Apeiron and run your first drift-detection experiment.
Every TOML section and key the config parser accepts, with defaults.
Detector classes, their options, and how detector output drives training.
The CL trainer, the updater modes, and what runs after drift is detected.
Suggested reading order¶
Installation and Quickstart — get a run going.
Architecture — how the pieces fit together at runtime.
Apeiron Configuration Reference — the required and optional configuration parameters.
Model Harness — how your model and stream loaders are exposed to the framework.
Drift Detectors — how monitoring decisions are made.
Continuous Learning — what happens after drift is detected.
Tracking Runs — sending run metrics to Weights & Biases or MLflow.
Contributing¶
Contribution guidelines, including our guidelines for AI/LLM-assisted contributions, are in CONTRIBUTING.md.
License¶
Apeiron is released under the Apache License 2.0.
Support¶
This work was supported by the U.S. Department of Energy (DOE), Office of Science, Office of Advanced Scientific Computing Research in alignment with DOE’s Genesis Mission.