Compositional Embeddings for Multi-Label One-Shot Learning