Scikit-Learn Integration 🧠

Classic ML pipelines made robust and reproducible.

What you'll learn

How to version and deploy sklearn models with automatic metadata extraction. Turn notebook scripts into production pipelines.

Why Scikit-Learn + flowyml?

• Auto-Extracted Metadata: Hyperparameters, feature importance, coefficients—all captured automatically.
• Pipeline Versioning: Version the entire preprocessing + model chain.
• Model Registry: Promote the best Random Forest to production.
• Easy Serving: Deploy sklearn models as APIs.

🎯 Model.from_sklearn() Convenience Method

The easiest way to create Model assets with full metadata extraction:

from sklearn.ensemble import RandomForestClassifier
from flowyml import Model

# Train your model
rf = RandomForestClassifier(n_estimators=100, max_depth=10)
rf.fit(X_train, y_train)

# 🎯 Use convenience method for full extraction
model_asset = Model.from_sklearn(
    rf,
    name="random_forest_classifier",
)

# Access auto-extracted properties
print(f"Framework: {model_asset.framework}")  # 'sklearn'
print(f"Class: {model_asset.metadata.properties.get('model_class')}")  # 'RandomForestClassifier'
print(f"Is Fitted: {model_asset.metadata.properties.get('is_fitted')}")  # True

# Hyperparameters auto-extracted!
print(f"Hyperparameters: {model_asset.hyperparameters}")
# {'n_estimators': 100, 'max_depth': 10, 'criterion': 'gini', ...}

# Feature importance (for tree-based models)
print(f"Has Feature Importance: {model_asset.metadata.properties.get('has_feature_importances')}")
print(f"Num Features: {model_asset.metadata.properties.get('num_features')}")

🧠 Pipeline Pattern

Return a sklearn.pipeline.Pipeline object for automatic serialization:

from sklearn.pipeline import Pipeline as SkPipeline
from sklearn.ensemble import RandomForestClassifier
from sklearn.preprocessing import StandardScaler
from flowyml import step, Model

@step
def build_pipeline():
    return SkPipeline([
        ('scaler', StandardScaler()),
        ('rf', RandomForestClassifier(n_estimators=100))
    ])

@step
def train(pipeline, X, y):
    pipeline.fit(X, y)

    # Return a Model asset with auto-extracted metadata
    return Model.from_sklearn(
        pipeline,
        name="trained_pipeline",
    )

🔧 Auto-Extracted Properties

The following properties are automatically extracted from sklearn models:

Property	Description
framework	Always 'sklearn'
model_class	Class name (RandomForestClassifier, etc.)
architecture	Same as model_class
hyperparameters	All model hyperparameters
is_fitted	Whether model has been fitted
has_feature_importances	True for tree-based models
num_features	Number of features (if fitted)
n_features_in	Number of input features
n_estimators	For ensemble models
num_estimators_fitted	Actual estimators fitted
max_depth	For tree-based models
num_classes	For classifiers
classes	Class labels (for classifiers)
coef_shape	Coefficient shape (for linear models)
intercept	Intercept (for linear models)

🌳 Supported Model Types

Full auto-extraction works for all sklearn estimators:

• Classifiers: RandomForest, GradientBoosting, SVM, LogisticRegression, etc.
• Regressors: RandomForest, Ridge, Lasso, ElasticNet, etc.
• Transformers: StandardScaler, PCA, etc.
• Ensembles: VotingClassifier, StackingClassifier, etc.
• Pipelines: sklearn.pipeline.Pipeline