The VotingClassifier is enhanced so that set_params can now be used to set a sub-estimator equals another object with fit and predict attribute. Example:

clf = VotingClassifier([('lr', LogisticRegression()), ('nb', GaussianNB())])
clf.set_params(nb=MultinomialNB()).fit(X, y)

An sub-estimator could also be set to None and it will not be used for fitting.
This enhancement enables gridsearch estimator types.

Our convention is not to raise validation errors in set_params, just as we generally do not validate in __init__. Validation is to be performed in fit.

I'd be interested to see how much you can reuse _BasePipeline if you moved that (or relevant parts of it) to sklearn.utils.metaestimators._BaseComposition (we could even consider making it public). Perhaps it should also be a CompositionMixin rather than a base class.

can you please try to rebase? You also have failing tests.