Saving and loading models

A trained model has no value if it exists only in the memory of a running process. The moment that process ends, the fit is lost. Serialisation converts the fitted object into bytes that can be written to disk, transferred to another machine, or loaded into a serving container — without repeating training.

Serialising with joblib

Sklearn's recommended serialiser is joblib, which handles numpy arrays efficiently (pickled numpy arrays are compact; joblib makes them more so):

import joblib
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression

# Assume pipe has been fitted
joblib.dump(pipe, "model.pkl")

Loading is symmetric:

loaded_pipe = joblib.load("model.pkl")
predictions = loaded_pipe.predict(X_new)

The loaded pipeline is an exact copy of the fitted one: all fitted attributes (coef_, mean_, scale_, etc.) are preserved. The training data is not stored — only the parameters derived from it.

What is serialised

joblib.dump serialises the Python object graph rooted at the pipeline. This includes:

• The pipeline structure (step names and step classes).
• All hyperparameters passed to constructors.
• All fitted attributes (the _ suffix attributes set during fit).

It does not include the training data, the Python interpreter, or any libraries. The loading environment must have compatible versions of sklearn, numpy, and joblib, or deserialization can fail. Pin your dependencies.

Verifying the load

After loading, confirm that predictions on a held-out sample match exactly:

import numpy as np
assert np.allclose(
    pipe.predict(X_test),
    loaded_pipe.predict(X_test)
), "Loaded model produces different predictions"
print("Verification passed.")

This is worth doing in a deployment pipeline as a smoke test — it catches version mismatches and file corruption before they reach users.

Versioning artefacts

A saved model without metadata is hard to trust. At minimum, store alongside the model file:

• The sklearn version (sklearn.__version__).
• The training dataset identifier or hash.
• The evaluation metrics from the held-out test set.
• The date of training.

A simple approach is to save a companion JSON file with these fields. More sophisticated ML platforms (MLflow, DVC) handle this automatically.

Where to go next

Next: the lab — build a complete pipeline with column-level preprocessing, evaluate it with cross-validation, then tune a hyperparameter with GridSearchCV.