Text this: Learning image by-parts using early and late fusion of auto-encoder features.