DataFrameMapper cannot leverage special fit_transform methods

[SandroCasagrande]

For transformers in a DataFrameMapper, it would be useful to be able to access native combined fit_transform methods. This would allow e.g. for the different semantics of PCA(..., copy=False).fit_transform(X) versus .fit(X).transform(X), or the optimized version of DictVectorizer.fit_transform for sparse matrices, etc.

The current implementation of DataFrameMapper can only access each transformers fit and transform method successively. It's implementaion of fit_transform is the default inherited from TransformerMixin which always calls fit(...).transform(...).

It should be quite simple to extend DataFrameMapper to account for this issue. I have the code prepared and will create a PR.

[FlorisHoogenboom]

It would be great to have this feature! There are many cases in which fit_transform should behave differently from fit().transform(), for examply when encoding categorical levels by leave-one-out means. Are there any plans to merge #152?

[dukebody]

Fixed in #169, which superseded #152 . Thanks!