Inference pipeline

fyt.core.inference_pipeline

Persistable inference pipeline: raw features in, predictions out.

Bundles the fitted preprocessing, feature-selection, target-encoding, and model objects from a training run into a single artifact that can be saved to disk, reloaded, and used for batch prediction without any training infrastructure (no MLflow, no dataset config).

InferenceMetadata

Bases: BaseModel

Provenance and interface metadata stored with a saved pipeline.

InferencePipeline

A fitted end-to-end pipeline for predicting on new, raw data.

The prediction flow mirrors training exactly: zero-imputation -> preprocessing transform -> feature selection transform -> model prediction -> inverse target encoding (classification).

Parameters:

Name	Type	Description	Default
pre_processor	Any	Fitted PreProcessor (or compatible transformer).	required
feature_selector	Any	Fitted FeatureSelector (or compatible transformer).	required
target_processor	Any	Fitted TargetProcessor (or compatible encoder).	required
model	Any	Fitted model exposing predict (and optionally predict_proba).	required
task	TaskType	Learning task type.	required
model_type	str	Model type identifier for provenance.	required

load(path) classmethod

Load a pipeline previously written by :meth:save.

Parameters:

Name	Type	Description	Default
path	str | Path	Path to the .joblib artifact.	required

Returns:

Type	Description
'InferencePipeline'	The deserialized InferencePipeline.

Raises:

Type	Description
FileNotFoundError	If the artifact does not exist.
TypeError	If the artifact is not an InferencePipeline.

predict(X)

Predict on raw feature data.

Parameters:

Name	Type	Description	Default
X	DataFrame	Raw features with the same columns used at training time (extra columns are tolerated if the preprocessor selects by name).	required

Returns:

Type	Description
Series	Predictions as a polars Series. For classification, labels are
Series	returned in the original (pre-encoding) space.

predict_proba(X)

Predict class probabilities on raw feature data.

Parameters:

Name	Type	Description	Default
X	DataFrame	Raw features with the training-time columns.	required

Returns:

Type	Description
DataFrame	One column per class, named by the original class labels when a
DataFrame	class mapping is available.

Raises:

Type	Description
ValueError	For regression tasks or models without predict_proba.

save(path)

Serialize the pipeline to disk with joblib.

Parameters:

Name	Type	Description	Default
path	str | Path	Destination file; a .joblib suffix is added if missing.	required

Returns:

Type	Description
Path	The final artifact path.