Learning Structured Output Representation using Deep Conditional Generative Models