Model Persistence

• Using Python's built-in persistence model (pickle)

from sklearn import svm
from sklearn import datasets

X,y = datasets.load_iris(return_X_y=True)
clf = svm.SVC().fit(X, y)

import pickle
s = pickle.dumps(clf)
clf2 = pickle.loads(s)
clf2.predict(X[0:1])

y[0]

0

• consider instead using joblib's dump & load methods.
• More efficient on large numpy arrays, but can only pickle to disk - not a string.

from joblib import dump,load
dump(clf, 'filename.joblib')

['filename.joblib']

clf = load('filename.joblib')

• Pickle (and joblib by extension) have maintainability & security issues. Never unpickle untrusted data.

• To rebuild a model with future versions of scikit-learn, save the following objects along with the pickled model:

  • training data (an immutable snapshot)
  • python source code
  • versions of scikit-learn and dependencies
  • cross-validation score obtained on the training data