A quick question about the definition of SimCLR

[HelloWorldLTY]

Hi, I intend to use InfoNCE loss in a graph, but I am confused about the definition of SimCLR. Based on the random sample approach mentioned in this paper, is it reasonable to choose only one pair in the dataset as positive pair while the rest of pairs are treated as negative pairs? What if all the pairs we sampled are actually positive or negative pairs? Thanks a lot.

[chentingpc]

positive pairs are generated from augmentations of the same image, and negative pairs are other (augmented) images in the mini-batch, so you're guaranteed with both positive and negative pairs

[HelloWorldLTY]

Got it, thanks a lot!