Stop 12 · Model Deployment & MLOpsTopics · 1/8

Saving & Loading Models (Pickle, joblib)

In short

A model that only lives inside your notebook dies when the kernel restarts. Saving it to a file is the first step to using it in an app, an API or a teammate's laptop.

Serializing a trained model to a file so it can be loaded and used later without retraining — Pickle is Python's general-purpose serializer, joblib is more efficient for models with large NumPy arrays.

Think of it like this

Making mango pickle

In summer, your nani cuts raw mangoes, adds spices and seals them in a jar. Months later anyone can open it and eat, no cutting or spicing needed. Saving a model is the same: all the training work, sealed in a file, ready to open later.

Watch it happen

From notebook to a separate app

Notebook

Disk

API process

fit(X, y)

1/4Train the whole pipeline, scaler and model together.

The big ideas

What you need to know

joblib for scikit-learn

joblib handles big NumPy arrays inside models more efficiently than plain pickle. Both use pickle's format underneath.

Save the whole pipeline

Save preprocessing and model together, so new data gets the exact same scaling and encoding.

Pin library versions

A model saved with one scikit-learn version may not load in another. Record versions in requirements.txt.

Only load trusted files

Loading a pickle can run hidden code. Never load model files from unknown sources.

See it in code

Walk through the code

python

1import joblib2from sklearn.datasets import load_iris3from sklearn.ensemble import RandomForestClassifier4from sklearn.pipeline import make_pipeline5from sklearn.preprocessing import StandardScaler6 7X, y = load_iris(return_X_y=True)8pipe = make_pipeline(StandardScaler(), RandomForestClassifier(random_state=0)).fit(X, y)9 10joblib.dump(pipe, "iris_model.joblib")       # save11 12loaded = joblib.load("iris_model.joblib")    # later, in another process13print(loaded.predict(X[:3]))

1/3Fit a pipeline so scaling travels with the model.

One line to save, one to load. Deep learning frameworks have their own formats, like model.save() in Keras.

Save yourself the pain

Common mistakes

Oops

Saving only the model and rebuilding the scaler by hand in the app.

Do this instead

Save the full Pipeline. Hand-rebuilt preprocessing almost always drifts from what training used.

Oops

Downloading a .pkl model from a random forum and loading it.

Do this instead

Treat pickle files like programs. Load only files you or your team created, or use a safer format such as skops.

Your mission

Save, restart, load

~15 min

Quick check

Why is it risky to load a pickle file from an unknown source?

Go deeper

1Save a trained model with joblib and reload it in a fresh script.
2Learn the security risk of loading pickled files from untrusted sources.
3Practice versioning your saved model files with a clear naming convention.