questions about contrastive-caption-traning details

[happywinder]

In contrastive caption fine-tuning ,how do you train on CC-3M?Does a contrastive learning on modified CC-3M(with help of shareCaptioner, the dense caption and original caption is regarded as positive)?

[Yif-Yang]

Yes, we used the enhanced CC3M dataset from ShareCaptioner, training it with one dense caption paired with the original caption.