kaiwenzha/rank-n-contrast

[NeurIPS 2023, Spotlight] Rank-N-Contrast: Learning Continuous Representations for Regression

rank-n-contrast's Issues

The effect of batch size

Hi, thank you for providing your repository. This is a great work!

Did you perform an ablation study on the batch size? I'm wondering if this method also works in a small batch size since I'm working on 3D data and I can't have a big batch size (in my case, 256 is considered big)

Thank you!

main_rnc.py

这个文件main_rnc.py 没有 val函数，此外，您的论文提到 MSE之外 R^2 等指标，也没有在代码中实现。数据集有好几个，目前只有agedb这个，想要一下作者完整版本code，谢谢~

实现跟跟论文图不一致

论文图的年龄回归问题，是把anchor的neighbor作为positive pair。但是在实现的时候，用的是anchor的random transform当作positive pair。。因此，最终的效果提升怎么排除是来自于图像数据增强呢？

Request: code for SkyFinder and UMAP visualization

Hello authors,

Thank you for the great work.

I am interested in the continuous feature pattern learned as visualized in the paper. Would it be possible to release code for processing the SkyFinder dataset and feature visualization?

Thanks for considering!