Hi, which version did you use for multi-positive text contrastive loss, and why it is pairwise circle loss or mpl which is proxy anchor loss, both losses are wildly different to Eq. 5 in the paper, can I realize the same performance with only naive implementation that exactly follows Eq. 5?
Hi, which version did you use for multi-positive text contrastive loss, and why it is pairwise circle loss or mpl which is proxy anchor loss, both losses are wildly different to Eq. 5 in the paper, can I realize the same performance with only naive implementation that exactly follows Eq. 5?