RankIQA.PyTorch icon indicating copy to clipboard operation
RankIQA.PyTorch copied to clipboard

Published 20 hours ago • verified publisher icon zheng-yuwei

rank loss 的 taget 似乎没有起作用？

Open yichengsu opened this issue 3 years ago • 1 comments

看上去target总是等于torch.ones(args.batch_size)，似乎没有起到作用？

https://github.com/zheng-yuwei/RankIQA.PyTorch/blob/fad8c07d7b45e59b1384cdb0a31724a395ee3a86/criterions/ranking_loss.py#L27-L28

yichengsu avatar Apr 06 '21 10:04 yichengsu

看上去target总是等于torch.ones(args.batch_size)，似乎没有起到作用？

https://github.com/zheng-yuwei/RankIQA.PyTorch/blob/fad8c07d7b45e59b1384cdb0a31724a395ee3a86/criterions/ranking_loss.py#L27-L28

前面一张图的分数比后面的高，第三个参数就应该是1，而一共有batch size对图，所以是torch.ones(args.batch_size)

STAR-811 avatar Aug 16 '21 12:08 STAR-811