The motivation is the same as PR #2561. In a nutshell, we want to be able to apply
the inverse_transform steps from a pipeline on an estimated parameter from the last step of the same pipeline. PR #2561 aims to achieve that goal by applying the inverse_transforms from all steps, except the last one if it misses an inverse_transform method. @GaelVaroquaux pointed out that it will fail in the cases where the last estimator implements both transform/inverse_transform methods and predict/score methods.

The alternative could be to implement a get_estimated method for Pipelines that explicitly retrieves a given attribute from the pipeline last step, and subsequently applies the inverse_transforms. The code would look like the following:

pca = PCA()
clf = LinearSVC()
pipeline = Pipeline([('pca', pca), ('clf', clf)])

pipeline.fit(X, y)
coef_ = pipeline.get_estimated('coef_')

assert X.shape[1] == coef_.shape[1]

Note that the final learner may itself be a Pipeline, a GridSearchCV, a MetaClassifier, or any valid combination in sklearn.

Any comments, suggestions?