[MRG+1] Fixed SelectKBest corner case: k=0

[griffinmyers]

Without this fix k = 0 will have the same behavior as k = 'all' based on how
the array indexing was written.

I've included a test that demonstrates proper behaviour.

[GaelVaroquaux]

LGTM. 👍 for merge. Thanks!

[coveralls]

Coverage remained the same when pulling f1e2e1c on griffinmyers:fix-select-k-best into eb10c4c on scikit-learn:master.

[jnothman]

This looks fine, but I wonder why an error would not be appropriate?

[griffinmyers]

In my case k=0 is a useful argument. Imagine you have a FeatureUnion of two Pipelines, each with its own selector. I'd like to be able to do a cross_val_score run where one of the pipelines produces zero features.

[jnothman]

Yes, I thought that might be the case. The ability to switch off a component in a FeatureUnion or Pipeline was mooted about a year ago (#1769).

But SelectPercentile(percentile=0) is explicitly supported, so I guess yours is a correct solution in any case.

[jnothman]

Thanks! merged