Multi label data layer

[kloudkl]

@leezl has implemented an MultiLabelDataLayer which is very important for many practical applications such as multi-label classification/annotation, detection, segementation and image retrieval.

@leezl, would you please kindly consider contributing your valuable works to the Caffe community?

[leezl]

Shoot. I was hoping no one would see that.

It's not done, still buggy, and I also was adding other features so it isn't as clean as it needs to be. It is also important to realize that a loss function appropriate to multilabel data needs to be implemented as well.

I'll try to clean it up and move my other changes into files that don't influence MultiLabelDataLayer. I also had to make changes to the Proto file so that the Datum could hold multiple labels, and to io.cpp, so it would load them. But the final layers have to be changed as well.

Multilabel learning is very experimental in my case.

[sergeyk]

Closing this due to alpha stage development.
@kloudkl and @leezl feel free to follow up at #149.

[kloudkl]

Thank @leezl for inspiring us! To resolve #88, I implemented a WARPLossLayer in #126 which a multilabel loss layer requiring multilabel data layer and that's why I asked for your help. I'm afraid there may be partial overlaps between those issues and #149.

You can open a pull request as soon as you start to separate the files according to the contribution protocol in #101.