Relative Margin for Contrastive Learning

We propose Relative Margin for multimodal contrastive learning that brings significant improvements to zero-shot image-text retrieval and image classification.