return coefficients from Pipeline object in sklearn

Question

I've fit a Pipeline object with RandomizedSearchCV

pipe_sgd = Pipeline([('scl', StandardScaler()),
                    ('clf', SGDClassifier(n_jobs=-1))])

param_dist_sgd = {'clf__loss': ['log'],
                 'clf__penalty': [None, 'l1', 'l2', 'elasticnet'],
                 'clf__alpha': np.linspace(0.15, 0.35),
                 'clf__n_iter': [3, 5, 7]}

sgd_randomized_pipe = RandomizedSearchCV(estimator = pipe_sgd, 
                                         param_distributions=param_dist_sgd, 
                                         cv=3, n_iter=30, n_jobs=-1)

sgd_randomized_pipe.fit(X_train, y_train)

I want to access the coef_ attribute of the best_estimator_ but I'm unable to do that. I've tried accessing coef_ with the code below.

sgd_randomized_pipe.best_estimator_.coef_

However I get the following AttributeError...

AttributeError: 'Pipeline' object has no attribute 'coef_'

The scikit-learn docs say that coef_ is an attribute of SGDClassifier, which is the class of my base_estimator_.

What am I doing wrong?

Ossieossietzky

Answer 1

You can always use the names you assigned to them while making the pipeline by using the named_steps dict.

scaler = sgd_randomized_pipe.best_estimator_.named_steps['scl']
classifier = sgd_randomized_pipe.best_estimator_.named_steps['clf']

and then access all the attributes like coef_, intercept_ etc. which are available to corresponding fitted estimator.

This is the formal attribute exposed by the Pipeline as specified in the documentation:

named_steps : dict

Read-only attribute to access any step parameter by user given name. Keys are step names and values are steps parameters.

Mcnamara

Answer 2

I think this should work:

sgd_randomized_pipe.named_steps['clf'].coef_

Raeraeann

Answer 3

I've found one way to do this is by chained indexing with the steps attribute...

sgd_randomized_pipe.best_estimator_.steps[1][1].coef_

Is this best practice, or is there another way?

Ossieossietzky

The named_steps method describe above is preferred – Farant

This worked well when using make_pipeline with many different classifiers! – Inextricable

Answer 4

In short, in scikit-learn there are two ways to access the estimators chained together in a Pipline: either retrieved by index or retrieved by name. (And each way again has two flavours, i.e. directly vs. indirectly.)

---

Firstly, as the User Guide of sklearn points out,

The Pipline is built using a list of (key, value) pairs (i.e. steps), where the key is a string containing the name you want to give this step and value is an estimator object.

Which indicates that:

1. a pipline is constructed by one or multiple estimator objects, in order. (just like a list)

   >>> from sklearn.pipeline import Pipeline
   >>> from sklearn.svm import SVC
   >>> from sklearn.decomposition import PCA
   >>> estimators = [('reduce_dim', PCA()), ('clf', SVC())]
   >>> pipe = Pipeline(estimators)
   >>> pipe
   Pipeline(steps=[('reduce_dim', PCA()), ('clf', SVC())])
   

2. and each estimator object has a name, either appointed by the user (with the key) or automatically set (e.g. by using make_pipeline utility function)

   >>> from sklearn.pipeline import make_pipeline
   >>> pipe = make_pipeline(PCA(), SVC())
   >>> pipe
   Pipeline(steps=[('pca', PCA()), ('svc', SVC())])
   

---

So finaly, we can access the estimators in a Pipline either

1. by indexing the Pipline:
   • directly through the Pipline object (just like a list)

     >>> pipe[0]
     PCA()
     >>> pipe[1]
     SVC()
     

   • indirectly through the steps attribute (actually a list of tuple)

     >>> pipe.steps
     [('pca', PCA()), ('svc', SVC())]
     >>> pipe.steps[0][1]
     PCA()
     >>> pipe.steps[1][1]
     SVC()
     

2. or by the name of steps/estimators:
   • directly through Pipline object (just like a dict or namedtyple)

     >>> pipe["pca"]
     PCA()
     >>> pipe["svc"]
     SVC()
     

   • indirectly through the named_steps attribute (actually a subclass of dict)

     >>> pipe.named_steps
     {'pca': PCA(), 'svc': SVC()}
     >>> pipe.named_steps["pca"]
     PCA()
     >>> pipe.named_steps["svc"]
     SVC()
     

From here on, I hope we could play around the piplines like a skilled plumber.

Elison