It would be nice to have an implementation of RFE.

some requirements:

• list-based feature dropping: user supplies [40, 20, 10] features means it would fit 4 rounds like [<all feats>, top 40, top 20, top 10]
• fixed-number-based feature dropping. e.g. if N features desired, fit in 4 rounds: can fit [<all feats>, top_n+40, top_n+20, top_n] if the fixed number is 20 features for fixed-number based dropping
• each round, rank top N features fitted. average all ranks from each round. return top N features
• use sklearn-style API
• support for doing it in CV