The output shape of VotingClassifier.transform is (n_classifiers, n_samples, n_classes), which is quite strange for sklearn and prevents it's use in pipelines (without extra work going through function transformer).
That kind of defeats the purpose of the transform imho.

At least n_samples should always be the first dimension, I think.
We could add a "flatten_transform" parameter or something so that transform returns (n_samples, n_classifiers * n_classes) so it can be used in a pipeline, without changing the current interface. Not sure if there is a good way to change the current default shape without major hassle.