Learning to Compose Domain-Specific Transformations for Data Augmentation