Accuracy

[Dadatata-JZ]

Hi there,

While I am walking through the code, I realize that for computing InfoNCE loss, the contrastive function creates pseudo labels, which is a list with all "0". However, this one is calculated as well when iterating the products of logits_and_labels for computing the accuracy in the train function. Should it be ignored when calculating the accuracy? Since all the labels are "0".

I may understand incorrectly. Thank you in advance and look forward to some helps.

[Dadatata-JZ]

Also, additionally, in the training processing, I realize that val_dataset and fg_val_dataset are essentially doing the same thing?! Can you clarify the necessity of validating twice?

Thank you in advance.