One question about AELoss

[FightStone]

Thanks for releasing the source code.
When i notice that in push aeloss, it subtracts a human ID number (num_tags), like this (82 rows in loss.py file): if self.loss_type == 'exp': diff = torch.pow(diff, 2) push = torch.exp(-diff) push = torch.sum(push) - num_tags What is the reason for this? Looking forward for your reply, thank you.

[738654805]

the pull and push loss really make me confued, they are brought from AE?

[bowenc0221]

Our AE loss is exactly the same as https://github.com/princeton-vl/pose-ae-train/tree/master/extensions/AE, except we implemented it in python.

[sathvikask0]

@FightStone after the step : push = torch.exp(-diff) the diagonal of the matrix all becomes one (exp(0) = 1 ), as on the diagonal we are subtracting the same tags. So when we add all the elements of the push matrix we need to remove numtags because that is exactly the number of elements on the diagonal then we get the actual loss value which is the sum of elements off-diagonal.