Subject-driven Text-to-Image Generation via Apprenticeship Learning