Could data augmentation improve the clustering?

[Jie2World]

SpCL is amazing project. You use Unified Contrastive Learning as loss function, which is different from SimCLR. But I find that the transformer don't use any data augmentation as SimCLR when I read the code of SpCl. I wonder whether the data augmentation (just like SimCLR or MoCo V2) could improve the clustering. For example, the centroids of data augmentation's feature outputs became a Cluster Centroids instead of Outlier Instance Features.

[Jie2World]

Sorry. I found transforms has done the work and got the meaning of Outlier Instance Features.