strange loss

[ZhuangLii]

HI,
RRID is an excellent work, I tried to apply the RRID to train on my own datasets, but I got strange loss curve , and it seems difficult to converge.

my config
lr : 0.01 lr_scheduler : multistep ID loss: arcface Triplet loss: semi hard optimizer: Ranger
could you please provide me some advise
THANKS for your time

[mohuang1996]

hi！
Did you reproduce the author's code? What is the map value of market1501 that you reproduced? I retrained according to the code given by the author and the result was lower than in the paper. The following is the result of my repetition:

Mean AP: 86.6%
CMC Scores market1501
top-1 94.2%
top-5 98.0%
top-10 98.7%

I don't know what went wrong.....

[ZhuangLii]

hi！
Did you reproduce the author's code? What is the map value of market1501 that you reproduced? I retrained according to the code given by the author and the result was lower than in the paper. The following is the result of my repetition:

Mean AP: 86.6%
CMC Scores market1501
top-1 94.2%
top-5 98.0%
top-10 98.7%

I don't know what went wrong.....

你是直接下git clone 复现的吗？我是在别的数据集上测试的，效果不好

[mohuang1996]

是的，我是想着先复现下作者提供的数据集在看看其他的。你复现了哪个数据集？效果如何呢？------------------&nbsp;原始邮件&nbsp;------------------ 发件人:&nbsp;&quot;ZhuangLii&quot;<[email protected]&gt; 发送时间:&nbsp;2020年11月3日(星期二) 晚上7:09 收件人:&nbsp;&quot;cvlab-yonsei/RRID&quot;<[email protected]&gt;; 抄送:&nbsp;&quot;mohuang1996&quot;<[email protected]&gt;;&quot;Comment&quot;<[email protected]&gt;; 主题:&nbsp;Re: [cvlab-yonsei/RRID] strange loss (#2)

[ZhuangLii]

是的，我是想着先复现下作者提供的数据集在看看其他的。你复现了哪个数据集？效果如何呢？------------------ 原始邮件 ------------------ 发件人: "ZhuangLii"<[email protected]> 发送时间: 2020年11月3日(星期二) 晚上7:09 收件人: "cvlab-yonsei/RRID"<[email protected]>; 抄送: "mohuang1996"<[email protected]>;"Comment"<[email protected]>; 主题: Re: [cvlab-yonsei/RRID] strange loss (#2)

是一个比赛，貌似这个不好调节