MLPipeline

class lightautoml.pipelines.ml.base.MLPipeline(ml_algos, force_calc=True, pre_selection=None, features_pipeline=None, post_selection=None)

Bases: object

Single ML pipeline.

Merge together stage of building ML model (every step, excluding model training, is optional):

• Pre selection: select features from input data. Performed by SelectionPipeline.

• Features generation: build new features from selected. Performed by FeaturesPipeline.

• Post selection: One more selection step - from created features. Performed by SelectionPipeline.

• Hyperparams optimization for one or multiple ML models. Performed by ParamsTuner.

• Train one or multiple ML models: Performed by MLAlgo. This step is the only required for at least 1 model.

Parameters:

• ml_algos (Sequence[Union[MLAlgo, Tuple[MLAlgo, ParamsTuner]]]) – Sequence of MLAlgo’s or Pair - (MlAlgo, ParamsTuner).

• force_calc (Union[bool, Sequence[bool]]) – Flag if single fold of ml_algo should be calculated anyway.

• pre_selection (Optional[SelectionPipeline]) – Initial feature selection. If None there is no initial selection.

• features_pipeline (Optional[FeaturesPipeline]) – Composition of feature transforms.

• post_selection (Optional[SelectionPipeline]) – Post feature selection. If None there is no post selection.

fit_predict(train_valid)

Fit on train/valid iterator and transform on validation part.

Parameters:

train_valid (TrainValidIterator) – Dataset iterator.

Return type:

LAMLDataset

Returns:

Dataset with predictions of all models.

predict(dataset)

Predict on new dataset.

Parameters:

dataset (LAMLDataset) – Dataset used for prediction.

Return type:

LAMLDataset

Returns:

Dataset with predictions of all trained models.

upd_model_names(prefix)

Update prefix pipeline models names.

Used to fit inside AutoML where multiple models with same names may be trained.

Parameters:

prefix (str) – New prefix name.

prune_algos(idx)

Prune model from pipeline.

Used to fit blender - some models may be excluded from final ensemble.

Parameters:

idx (Sequence[int]) – Selected algos.