Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

评价指标 #2

Open
SuQinghang opened this issue May 15, 2021 · 2 comments
Open

评价指标 #2

SuQinghang opened this issue May 15, 2021 · 2 comments

Comments

@SuQinghang
Copy link

SuQinghang commented May 15, 2021

您好,我最近在阅读论文和复现Multi-scale learning baseline时对论文中所采用的评价指标有几点迷惑,还望解答!
首先说一下我的实验设置:

  1. 数据组织按照您所公布的源代码,backbone采用resnet34(imagenet pretrained)。
  2. 我目前只在Multi-Similarity Loss上进行了测试。因为每个batch中fine、middle、coarse的图像各占1/3,在某个scale下进行损失计算时,我只选择当前scale的1/3图像作为anchor来组织样本对。如假设batch-size为1,那么一个batch中有K*3张图片,前K张为fine-scale下采样出来的图片,那么在fine上进行损失函数计算时,我只使用K张图片作为anchor,negative的话则在整个bath上选择。
  3. 训练时batch-size我设置的为5, MS Loss的学习率等参数按照作者源代码中进行设置。
  4. 评价指标目前只使用了Recall@K,与MS Loss中相同。计算方式为3个尺度分别计算Recall@K,求平均。

下面说一下我遇到的问题:

  1. 目前只在Vechicle上进行测试,无法获得论文中MS Loss下的实验效果(我获得的Recall@1, Recall@10都要高。
  2. 评价指标ASI中需要指定k,论文中公布的ASI指标是在k为多少时得到的。

感谢您的查看,期待您的指教和解答!

@SupetZYK
Copy link
Owner

实验设置
1.对animal,是没有pretrain的,对vehicle和product是有pretrain的
2.对一个batch中的3*K张图像,为了计算loss方便,我们是分开计算的,也就是fine部分用前K个,middle部分用中间K个,coarse部分用后K个
回复问题
1.正如你提到的,你用到的负样本数量比我们在paper中设置的多,相当于增加了batch size,这确实是会提升recall的。
2.ASI使用的各个K下的平均值。

@SuQinghang
Copy link
Author

您的解答对我有极大帮助,非常感谢!祝好!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants