[ENH] Pipeline partial transform

[elliot-hallmark]

It's helpful to evaluate a pipeline up to a certain point and then inspect the output. In Caffe this is achieved with the kwarg end. In my proposal if a pipeline did augmentation, feature interaction, scaling, anova feature selection, and then classification, you could run pipeline.transform(X, end='anova') and get the transformed data prior to classification.

This if helpful in investigating the state of data deeper into the pipeline in order to think how to improve your pipeline

[glemaitre]

We are inheriting from scikit-learn and thus the feature should be proposed upstream. Be aware that you can do what you want with indexing:

# call transform on all steps but the last one
pipeline[:-1].transform(X)