Question about logits_mask

[Alva-2020]

Thanks very much for your contributions. But I have some trouble when I read the code and paper. ` # mask-out self-contrast cases logits_mask = torch.scatter( torch.ones_like(mask), 1, torch.arange(batch_size * anchor_count).view(-1, 1).to(device), 0 )

    mask = mask * logits_mask

    # compute log_prob
    exp_logits = torch.exp(logits) * logits_mask
    #
    log_prob = logits - torch.log(exp_logits.sum(1, keepdim=True))`

In the above code, is the logits_mask used to filter out the negative pairs? Why is it a square matrix whose diagonal is 0 and 1 of other elements? I think the logits_mask = ~ mask, which is to flag which pair is negative? Thank you for your reply.

[QishengL]

0 because you don't want to use the inner product of itself. 1 means other positions are good to use. After mask * logits_mask. You get all the positions that are inner products of positive samples and exclude the sample itself.

What I do not understand is log_prob = logits - torch.log(exp_logits.sum(1, keepdim=True)). Why do we use logits to minus the log of the denominator of the equation in the paper? Did you figure it out?

[fanchi]

log(a/b) = log(a) - log(b)

[shuaiNJU]

0 because you don't want to use the inner product of itself. 1 means other positions are good to use. After mask * logits_mask. You get all the positions that are inner products of positive samples and exclude the sample itself.

What I do not understand is log_prob = logits - torch.log(exp_logits.sum(1, keepdim=True)). Why do we use logits to minus the log of the denominator of the equation in the paper? Did you figure it out?

hello，you said that "After mask * logits_mask. You get all the positions that are inner products of positive samples and exclude the sample itself.", however, each row of the mask matrix is a one-hot coding vector, that's means only an augmentation sample that is the same origin as anchor is used as positive, and no other augmentation samples that belong to the same label as anchor are used. So can you tell me which code shows other augmentation samples that belong to the same label as anchor are used as positives? Thanks a lot!

[Arsiuuu]

0 because you don't want to use the inner product of itself. 1 means other positions are good to use. After mask * logits_mask. You get all the positions that are inner products of positive samples and exclude the sample itself. What I do not understand is log_prob = logits - torch.log(exp_logits.sum(1, keepdim=True)). Why do we use logits to minus the log of the denominator of the equation in the paper? Did you figure it out?

hello，you said that "After mask * logits_mask. You get all the positions that are inner products of positive samples and exclude the sample itself.", however, each row of the mask matrix is a one-hot coding vector, that's means only an augmentation sample that is the same origin as anchor is used as positive, and no other augmentation samples that belong to the same label as anchor are used. So can you tell me which code shows other augmentation samples that belong to the same label as anchor are used as positives? Thanks a lot!

Did you resolve it? Looking forward to your reply.

[jlliRUC]

hello，you said that "After mask * logits_mask. You get all the positions that are inner products of positive samples and exclude the sample itself.", however, each row of the mask matrix is a one-hot coding vector, that's means only an augmentation sample that is the same origin as anchor is used as positive, and no other augmentation samples that belong to the same label as anchor are used. So can you tell me which code shows other augmentation samples that belong to the same label as anchor are used as positives? Thanks a lot!

For multi augmentions in unsupervised scenario or multi samples with same labels in supervised scenario, the mask matrix is not one-hot vector. It contains the label info of all positives.