FeatureUnion provided no simple way to tell which parts of the stack belonged to which transformer. This provides a feature_ptr_ attribute (of the form of csr_matrix.indptr) to solve that.

Caveats:

• it can only be determined when fit_transform, not fit, is called
• it will be incorrect if one of the sub-transformers has a change of parameters that affects the output size at transform time without a refit.

Both these caveats would be solved if each transformer in sklearn provided a way to get the number of output features. I suggest a transformed_width_ attribute or get_transformed_width() method (better name?), the latter making it more clear that the output can be affected by set_params. I also suggest this be posed as an Easy Issue for someone to tackle.

Finally, feature_ptr_ is compact and versatile, but it might be more usable if I add a method to get this data as a dict from transformer-name to slices.