Saving & reloading a pipeline? #32652

Yes — when you pickle the Pipeline, the entire pipeline is saved, including all steps (StandardScaler, GaussianNB, etc.) and their fitted parameters.

When you load it back, you can use it directly on new data:

with open('my_model', 'rb') as file: pipe = pickle.load(file)

y_pred = pipe.predict(X_new)

You do not need to manually reapply StandardScaler. The loaded pipeline will automatically apply preprocessing → model prediction, exactly as it did before saving.

Just note: scikit-learn recommends saving models using joblib instead of pickle:

from joblib import dump, load dump(pipe, 'my_model.joblib') pipe = load('my_model.joblib')

Hi!

When you serialize a scikit-learn Pipeline using pickle (or joblib), the saved file contains the entire state of the pipeline, including the fitted parameters of all steps.

This means that:

1. Yes, the StandardScaler is fully saved as part of the pipeline.
2. No, you do not need to manually apply the scaling function to new data.

When you load the pickled pipeline back and call predict(), the pipeline automatically passes the new data through the pre-fitted StandardScaler (reusing the mean and variance calculated during fit()) before handing the scaled data to the GaussianNB classifier.

Example

import pickle
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.naive_bayes import GaussianNB

# Fit and save
pipe = make_pipeline(StandardScaler(), GaussianNB())
pipe.fit(X_train, y_train)

with open('my_model.pkl', 'wb') as file:
    pickle.dump(pipe, file)

# Later... Load and predict directly
with open('my_model.pkl', 'rb') as file:
    loaded_pipe = pickle.load(file)

# The loaded_pipe will automatically scale X_test using the scale/mean of X_train!
predictions = loaded_pipe.predict(X_test)

For large scikit-learn models (especially those containing numpy arrays like random forests), it is highly recommended to use joblib instead of pickle as it is much more efficient at handling large numerical arrays.

import joblib
joblib.dump(pipe, 'my_model.joblib')
loaded_pipe = joblib.load('my_model.joblib')

Hi!

When you serialize a scikit-learn Pipeline using pickle (or joblib), the saved file contains the entire state of the pipeline, including the fitted parameters of all steps.

This means that:

1. Yes, the StandardScaler is fully saved as part of the pipeline.
2. No, you do not need to manually apply the scaling function to new data.

When you load the pickled pipeline back and call predict(), the pipeline automatically passes the new data through the pre-fitted StandardScaler (reusing the mean and variance calculated during fit()) before handing the scaled data to the GaussianNB classifier.

Example

import pickle
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.naive_bayes import GaussianNB

# Fit and save
pipe = make_pipeline(StandardScaler(), GaussianNB())
pipe.fit(X_train, y_train)

with open('my_model.pkl', 'wb') as file:
    pickle.dump(pipe, file)

# Later... Load and predict directly
with open('my_model.pkl', 'rb') as file:
    loaded_pipe = pickle.load(file)

# The loaded_pipe will automatically scale X_test using the scale/mean of X_train!
predictions = loaded_pipe.predict(X_test)

For large scikit-learn models (especially those containing numpy arrays like random forests), it is highly recommended to use joblib instead of pickle as it is much more efficient at handling large numerical arrays.

import joblib
joblib.dump(pipe, 'my_model.joblib')
loaded_pipe = joblib.load('my_model.joblib')

Yes, the fitted Pipeline is serialized as one estimator, including the StandardScaler, its learned mean/scale, and the classifier. After loading it, call predict(X_new) directly; do not scale X_new manually.

For example:

from joblib import dump, load

dump(pipe, "model.joblib")
loaded = load("model.joblib")
predictions = loaded.predict(X_new)

Two practical caveats: only load pickle/joblib files from trusted sources because deserialization can execute arbitrary code, and keep the scikit-learn and dependency versions compatible with the environment used for training.