Efficient Hierarchical Embedding for Learning Coherent Visual Styles