question about emb loss

[256785]

trafficstars

def call(self, q_reps, p_reps): if self.negatives_cross_device: # This gathers both negatives and positives. # It could likely be optimized by only gathering negatives. q_reps = self._dist_gather_tensor(q_reps) p_reps = self._dist_gather_tensor(p_reps) scores = self.compute_similarity(q_reps, p_reps) / self.temperature scores = scores.view(q_reps.size(0), -1)

    target = torch.arange(scores.size(0), device=scores.device, dtype=torch.long)
    target *= (p_reps.size(0) // q_reps.size(0))
    return self.cross_entropy(scores, target)

in the code，does it use ContrastiveLoss following the paper？

[Muennighoff]

yes thats contrastive loss

[256785]

target = torch.arange(scores.size(0), device=scores.device, dtype=torch.long) target *= (p_reps.size(0) // q_reps.size(0))

why use target as this way，a little confuse

[256785]

I have some sense each query refers to some samples，use divided to count the num of samples，and use arrange with multiply to find the positive item index. Maybe is that?