limen.sfd.reference_architecture
Model-function implementations and templates that own the training logic behind each SFD.
Canonical docs
What this package owns
Owns the reference model classes (and their function forms) that implement fit/predict for each decoder family, without the packaged data pipeline.
Does not own the experiment shape or manifest (owned by the matching limen.sfd.foundational_sfd module) or experiment execution.
Key entry points
| Entry point | Use case | Notes |
|---|---|---|
ReferenceModel | Base class for reference models | Shared fit/predict contract |
LogRegBinary / logreg_binary | Logistic-regression binary model | Class and function forms |
DLinearRegressor / dlinear_regressor | Linear time-series regressor | Class and function forms |
LightGBMBinary / lightgbm_binary | LightGBM binary classifier | Class and function forms |
XGBoostRegressor / xgboost_regressor | XGBoost regressor | Class and function forms |
RandomBinary / random_binary | Baseline random classifier | Class and function forms |
RuleBasedStrategy / rule_based | Predicate-driven, non-learned strategy | Uses limen.sfd.rule_based |
TabPFNBinary / tabpfn_binary | TabPFN binary classifier | Optional; requires tabpfn |
Adjacent modules
limen.sfd.foundational_sfdpairs each of these models with an experiment shape and manifest.limen.inference.Sensorwraps aReferenceModelfor live bar-by-bar prediction.limen.metricsis commonly used inside these model functions to score outputs.
Quick orientation
reference_architecture/
├── base.py # ReferenceModel base class
├── logreg_binary.py
├── dlinear_regressor.py
├── lightgbm_binary.py
├── xgboost_regressor.py
├── random_binary.py
├── rule_based.py
└── tabpfn_binary.py # Optional
Things to know
- Each module exposes both a class (e.g.
LogRegBinary) and a function form (e.g.logreg_binary); pick the class for custom subclassing and the function for manifest wiring. - These models are the template for custom SFD work: subclass
ReferenceModelor copy a module to start. - A reference architecture owns the training logic; the matching foundational SFD owns the manifest and data pipeline.