I know there are several out of the box methods of saving the model.

However I want to save the model as a database blob.

I've seen examples where people extract 'coefs' etc from the dict of the estimator but when I try this with RandomForrestRegressor (for example) it says there are no coefs.

I've seen yet other examples that claim all estimators have a 'save' method somewhere but I can't get that to work.

Is there any way of getting the model data from the fitted estimator in a way that will work universally for all estimators?

I would then base64 the model data and persist it to my Database for later use. There seems little documentation on this aside from just 'use pickle or joblib' etc.

I know there are several out of the box methods of saving the model.

However I want to save the model as a database blob.

I've seen examples where people extract 'coefs' etc from the dict of the estimator but when I try this with RandomForrestRegressor (for example) it says there are no coefs.

I've seen yet other examples that claim all estimators have a 'save' method somewhere but I can't get that to work.

Is there any way of getting the model data from the fitted estimator in a way that will work universally for all estimators?

I would then base64 the model data and persist it to my Database for later use. There seems little documentation on this aside from just 'use pickle or joblib' etc.

• scikit-learn
• persistence

1 Answer 1

This is a good summary of some methods of serializing sklearn models.

pickle works on most python objects and I see that it works fine for sklearn models too. For example

def model_to_base64(model):
    import pickle, base64
    return base64.b64encode(pickle.dumps(model)).decode('utf-8')

def base64_to_model(encoded_model):
    import pickle, base64
    return pickle.loads(base64.b64decode(encoded_model))

Tested with several kinds of estimators

from sklearn.ensemble import RandomForestClassifier, RandomForestRegressor
from sklearn.linear_model import LogisticRegression, LinearRegression
from sklearn.svm import SVC, SVR
from sklearn.cluster import KMeans
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline
from sklearn.tree import DecisionTreeClassifier

models = [
    RandomForestClassifier(n_estimators=10, random_state=42).fit([[1, 2], [3, 4], [5, 6]], [0, 1, 0]),
    RandomForestRegressor(n_estimators=10, random_state=42).fit([[1, 2], [3, 4], [5, 6]], [10, 20, 30]),
    LogisticRegression().fit([[1, 2], [3, 4], [5, 6]], [0, 1, 0]),
    LinearRegression().fit([[1, 2], [3, 4], [5, 6]], [10, 20, 30]),
    SVC(probability=True).fit([[1, 2], [3, 4], [5, 6]], [0, 1, 0]),
    SVR().fit([[1, 2], [3, 4], [5, 6]], [10, 20, 30]),
    KMeans(n_clusters=2, random_state=42).fit([[1, 2], [3, 4], [5, 6]]),
    DecisionTreeClassifier().fit([[1, 2], [3, 4], [5, 6]], [0, 1, 0]),
    Pipeline([
        ('scaler', StandardScaler()),
        ('classifier', LogisticRegression())
    ]).fit([[1, 2], [3, 4], [5, 6]], [0, 1, 0]),
]

for i, model in enumerate(models):
    encoded_model = model_to_base64(model)
    print(f"Model {i + 1} Serialized: {encoded_model[:100]}...")
    
    restored_model = base64_to_model(encoded_model)
    print(f"Model {i + 1} Restored: {restored_model.__class__.__name__}")
    
    assert isinstance(restored_model, model.__class__)