the MAP on NUSWIDE is just about 0.685 when the bits length is 48

[wzb1005]

Hi, I have tried the code, but the MAP on NUSWIDE is just about 0.685 when the bits length is 48, and it is far from the reported result 0.7218. So why? Yours sincerely.

[caozhangjie]

Have you tried our pretrained evaluation model?

[wzb1005]

I have tried your provided caffemodel "nuswide_81_48.caffemodel", but the MAP is just "0.69239" . I have provided the "deploy.prototxt" as an attachment for you to check. I'm looking forward to your responding. ( If the "Dropout" layer should be deleted??) Thank you! Yours Sincerely. WZB.

deploy.txt (Just change the file type to ".prototxt")

------------------ 原始邮件 ------------------ 发件人: "Zhangjie Cao"[email protected]; 发送时间: 2019年8月9日(星期五) 凌晨1:26 收件人: "thuml/DPH"[email protected]; 抄送: "ww"[email protected];"Author"[email protected]; 主题: Re: [thuml/DPH] the MAP on NUSWIDE is just about 0.685 when the bitslength is 48 (#2)

Have you tried our pretrained evaluation model?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[caozhangjie]

https://github.com/thuml/DPH/blob/master/models/train/nuswide_81/train_val.prototxt The train prototxt is here.

[wzb1005]

Thank you, I know this file. And what I used is just this file. I have found that : the "PairwiseLossLayer" and "QuantizationLoss" have some "threshold" fields (in the "train_val.prototxt" file) that don't described in the paper. I doubt if the "prototxt" file is right or not.