| name | integrate-apeiron |
| description | Add apeiron's continual-learning / drift-detection capabilities to a user's
EXISTING training framework. Use when the user already has their own training
loop (vanilla PyTorch, Lightning, HF Trainer, etc.) and wants to bolt on drift
detection and/or CL adaptation rather than adopt apeiron's runner. Inspects the
user's repo, recommends the lightest viable integration path, writes the
adapter glue into their repo, and smoke-tests it. Assumes apeiron is importable
(`import apeiron`) — if not, run install-apeiron first. For a self-contained
apeiron run on custom data, use custom-experiment instead.
|
| argument-hint | [target_project_dir] |
| user-invocable | true |
| allowed-tools | ["Bash","Read","Write","Edit","Glob","Grep"] |
Integrate apeiron into the user's existing training framework, lightest path first.
Arguments
$1: The user's project directory. Defaults to the current working directory.
Background: what apeiron exposes (verify against source, don't assume)
- Drift detectors are standalone —
detector.update(metric_value: float) -> DriftSignal; the signal carries drift_detected, regime (LearningRegime), and drift_score. This is the lowest-coupling entry point. Build one via from apeiron.drift_detection import ADWINDetector (or the others).
- ContinuousMonitor drives the full loop but requires a
BaseModelHarness wrapping the model + data stream, plus a Config and a detector. Mirror src/main.py for exact construction.
- CL updaters (EWC/JVP/KFAC) are harness-coupled — they take a
modelHarness, so using them implies the harness/monitor route.
Read these before writing any glue so it matches the current API:
cat src/main.py
cat src/apeiron/drift_detection/detectors/base.py
cat src/apeiron/drift_detection/load_drift_detector.py
grep -nA12 "class ContinuousMonitor" src/apeiron/driver/continuous_monitor.py
Procedure
1. Confirm apeiron is importable
python -c "import apeiron; print('apeiron', apeiron.__file__)" 2>&1
If this fails, stop and direct the user to the install-apeiron skill, then resume.
2. Discover the user's framework (detect at runtime)
In $1, find the training loop and the evaluation signal:
- Detect the stack:
grep -rlE "pytorch_lightning|lightning|transformers|Trainer|accelerate" $1 and look for a manual loop (loss.backward(), optimizer.step()).
- Locate where a scalar quality metric per step/epoch is available (val accuracy, loss) — this is what a detector consumes.
- Locate the model object and the data iterator (needed only if the full path is chosen).
Summarize what you found before proposing anything.
3. Recommend the lightest path that meets the need, and confirm
Based on what the user wants out of apeiron:
- Just detect drift / trigger their own retrain → detectors-only (no harness). Lowest coupling — recommend this unless they need apeiron's CL math.
- Want apeiron's CL regularizers (EWC/JVP/KFAC) or the full monitor→adapt loop → harness + ContinuousMonitor.
Present the recommendation with its trade-offs and get the user's pick before writing code.
4a. Detectors-only adapter (lightest)
Write a small module into the user's repo (e.g. <their_pkg>/apeiron_drift.py) that:
- Constructs a detector once (choice of ADWIN/KSWIN/PageHinkley).
- Exposes a hook called from their existing eval step:
signal = detector.update(metric); if signal.drift_detected: <their callback>.
- Leaves the decision of what to do on drift (log / retrain / reload) to a user-provided callback.
Wire the hook into their loop with a minimal, clearly-marked edit.
4b. Harness + monitor adapter (full)
When CL adaptation is wanted:
- Write a
BaseModelHarness subclass in their repo that wraps their existing model and data loaders, implementing the abstract methods (read src/apeiron/model/torch_model_harness.py and examples/mnist/model.py for the current set — includes get_optmizer, update_data_stream, get_stream_dataloader, get_hist_dataloaders, get_train_dataloaders, get_criterion).
- Build a
Config (via build_config from a small TOML, or constructed directly) selecting the detector and continual_learning.update_mode.
- Construct and run
ContinuousMonitor exactly as src/main.py does.
5. Smoke-test the integration
Prove the wiring with a tiny run before handing back:
- Detectors-only: a short script feeding a handful of synthetic metric values through the hook, asserting a
DriftSignal comes back and the drift callback fires on an obvious shift.
- Full path: run their loop (or the monitor) for a couple of iterations with minimal settings (small
max_iter, few stream updates, device=cpu, logging.backend=none).
Fix and re-run until it completes cleanly.
6. Report
Summarize: detected stack, chosen path, files added/edited (with the exact insertion points), how to invoke it, and what happens on drift. Note any assumptions the user should revisit (e.g. which metric drives detection, detector sensitivity params).
Notes
- Keep edits to the user's training loop minimal and clearly commented so they remain easy to revert.
- Do not assume detector/monitor/harness signatures — they are read from source in this skill precisely so the glue doesn't rot.
