xyupeng / ContrastiveCrop

[CVPR 2022 Oral] Crafting Better Contrastive Views for Siamese Representation Learning
MIT License
284 stars 27 forks source link

More models and datasets coming soon? #19

Closed DWCTOD closed 1 year ago

DWCTOD commented 1 year ago

你好,请问后续会更新模型的权重和数据集吗? 另外想请问一下大佬,关于对比学习,如何去评价模型embedding是否学的“够好”(表征能力)?好像一般是根据下游任务去评价的,但是有些任务很难去定义这样子的任务。例如业务场景是那种社媒的?

xyupeng commented 1 year ago

后续更新可能还需要一段时间。表征能力的评估在cv领域一般有分类linear probing,检测/分割fine-tune, partial fine-tune, freeze backbone等等。一般来说预训练的表征够好,则在下游任务上的初期收敛速度较快。

DWCTOD commented 1 year ago

谢谢大佬的解答,感谢