Describing Like Humans: On Diversity in Image Captioning