Persistence

Saving and loading a result

result.save("runs/experiment_01")

Creates a directory with:

File	Contents
step1_fold_df.parquet	Per-fold metrics, step 1
step3_fold_df.parquet	Per-fold metrics, step 3
step4_fold_df.parquet	Per-fold metrics, step 4
features.npy + metadata	Full feature matrix
features_reduced.npy + metadata	Reduced feature matrix
selector.pkl	Fitted FeatureSelector
step1_cv_result.pkl	Raw CVResult list, step 1
step3_cv_result.pkl	Raw CVResult list, step 3
step4_cv_result.pkl	CVResult, step 4
meta.json	Scalar fields (model name, stage, params, …)

Reload with:

from h2ml import PipelineResult

result = PipelineResult.load("runs/experiment_01")

Warning

The .pkl files use joblib (pickle under the hood). Only load results from trusted sources — pickle executes arbitrary code on deserialisation.

Resuming a partial run

Save after any partial run and resume later:

# Save after steps 1–2
result = pipeline.run_step1_to_step2(store)
result.save("runs/partial")

# Reload and continue from step 3
result = PipelineResult.load("runs/partial")
result = pipeline.run_from_step3(result)

Saving the final model

FinalModel is saved independently of PipelineResult — it is the deployment artifact:

from h2ml.pipeline.final_model import FinalModel

final = result.build_final_model()
final.save("models/final.pkl")

final = FinalModel.load("models/final.pkl")
final.predict(X_new)