TopolearnOnce you know the one API, you can use any of the library's hundreds of models and transformers.
scikit-learn is built around one small interface. An estimator is any object with a fit method. You configure it with keyword arguments in the constructor (the hyperparameters), then learn from data with fit(X, y).
Estimators come in a few flavours that add one more method:
- Predictors (classifiers, regressors) add
predict(X), and usuallyscore(X, y). Classifiers often addpredict_proba(X). - Transformers (scalers, encoders, PCA) add
transform(X), andfit_transform(X)as a shortcut for fitting then transforming the same data. - Clusterers add
fit_predictor alabels_attribute.
Two conventions make the API predictable. Constructor arguments are stored unchanged, so model.C is what you passed. Everything learned from data is stored in attributes ending in an underscore, such as coef_, classes_ or mean_, and only exists after fit. Calling predict before fit raises NotFittedError.
fit returns the estimator itself, so model = LogisticRegression().fit(X, y) is a common one-liner. Calling fit again throws away what was learned and starts over; it does not continue training.
Because the interface is the same everywhere, swapping a logistic regression for a random forest is usually a one-line change, and generic tools such as Pipeline, cross_val_score and GridSearchCV work with any estimator.
Example
import numpy as np
from sklearn.linear_model import LinearRegression
from sklearn.preprocessing import StandardScaler
X = np.array([[1.0], [2.0], [3.0], [4.0]])
y = np.array([3.0, 5.0, 7.0, 9.0])
model = LinearRegression() # hyperparameters go in the constructor
model.fit(X, y) # learns from data, returns self
print(model.coef_, model.intercept_) # [2.] 1.0 (learned attributes end in _)
print(model.predict([[10.0]])) # [21.]
scaler = StandardScaler()
print(scaler.fit_transform(X).ravel()) # [-1.34164079 -0.4472136 0.4472136 1.34164079]
print(scaler.mean_) # [2.5]Remember
- Configure in the constructor, learn with
fit, thenpredictortransform - Learned state lives in attributes with a trailing underscore
- Calling
fitagain retrains from scratch
Watch out for
- Passing training data to the constructor instead of to
fit.