TRACE: Transform Aggregate and Compose Visiolinguistic Representations for Image Search with Text Feedback