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

lfw准确率为99.73是如何得到的?使用的什么数据集! #21

Closed
jianbaba opened this issue Aug 9, 2018 · 12 comments
Closed

Comments

@jianbaba
Copy link

jianbaba commented Aug 9, 2018

您好,我有看您的实验都是使用的64层spereface结构进行训练的!那么请问,您有在别的结构上训练过么?我看arcface实用您的损失,arcface的结构能够达到99.7~99.8(仅适用vgg2或者ms)

@happynear
Copy link
Owner

我用的是20层的spherenet哎...目前用vgg2可以达到99.55%,估计64层能到99.7%吧。

@jianbaba
Copy link
Author

jianbaba commented Aug 9, 2018

@happynear 我是将am和cosface一起看的,这两篇的方法是一样的吧?还有就是我使用你给的网络结构去训练,相同的学习率和衰减方式,为什么准确率只有60多?我用的是face_train_test_auto.prototxt!!!您release的几个有什么差别么?

@happynear
Copy link
Owner

恩一样的。
你先别用auto,那个算法不成熟。你先用https://github.com/happynear/AMSoftmax/blob/master/prototxt/face_train_test.prototxt 试试。

@jianbaba
Copy link
Author

jianbaba commented Aug 9, 2018

@happynear 好的,还有就是您在casia和vgg2上都是用的一样的学习率,和衰减方式么?( 0.1 and is divided by 10 at the 16K, 24K and 28K iterations. The training is finished at 30K iterations )一共用了几个gpu呢?如果我在64层做实验,请问您有什么建议么?还是这样的学习率和衰减方式么?

@happynear
Copy link
Owner

vgg2上要把这些step扩大一倍。不管用几个gpu,保持batch size为256就行。

@jianbaba
Copy link
Author

jianbaba commented Aug 9, 2018

@happynear 好的,谢谢老哥!

@jianbaba
Copy link
Author

jianbaba commented Aug 9, 2018

@happynear 对了,测试的时候deploy选用这个face_deploy_mirror_normalize.prototxt??还是直接把face_train_test.prototxt 改
(最后一层为?)
layer {
name: "norm1"
type: "Normalize"
bottom: "fc5"
top: "norm1"
}

@happynear
Copy link
Owner

face_deploy_mirror_normalize.prototxt这个比较好

@jianbaba
Copy link
Author

jianbaba commented Aug 9, 2018

@happynear 请问您推荐的这个prototxt原理是什么?图片为什么可以contact一起送入网络(是变成了6通道么?)之后为什么还能分开?
相比sphereface是将图片以及翻转图片分别送入网络,得到的特征进行串联!

@happynear
Copy link
Owner

https://github.com/happynear/NormFace/blob/master/MirrorFace.md
看这个。也是翻转图片。

@jianbaba
Copy link
Author

jianbaba commented Aug 9, 2018

@happynear 所以这个操作是代替了sphereface的做法是么?所以你这个最后不需要再去翻转后再输入到有镜像的网络了是么?直接用512维特征进行运算,而不是1024维度?

@happynear
Copy link
Owner

恩。

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