-
Notifications
You must be signed in to change notification settings - Fork 143
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
about train GCNV #60
Comments
@Linsongrong 您好,GCN-V的目标是让每个vertex只输出一个confidence,用来表示其属于特别类别的置信度,故而nclass设置为1 |
@yl-1993 @Linsongrong 你好,检测了下GCN-V输入的是feature_dim=256,但我们提取的特征集都是512,而看你们都是提取512维特征,这个是不是相互矛盾呢?向各位求解 |
提的是256维啊
获取 Outlook for Android<https://aka.ms/ghei36>
…________________________________
From: cvlmm <[email protected]>
Sent: Tuesday, September 15, 2020 11:36:50 AM
To: yl-1993/learn-to-cluster <[email protected]>
Cc: Linsongrong <[email protected]>; Mention <[email protected]>
Subject: Re: [yl-1993/learn-to-cluster] about train GCNV (#60)
@yl-1993<https://github.com/yl-1993> @Linsongrong<https://github.com/Linsongrong> 你好,检测了下GCN-V输入的是feature_dim=256,但我们提取的特征集都是512,而看你们都是提取512维特征,这个是不是相互矛盾呢?向各位求解
―
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub<#60 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AHFLQIE43MJFDQSVCXYUZMDSF3ONFANCNFSM4PKGIX5Q>.
|
@yl-1993 @Linsongrong ,再次请教你们 |
你好,不知道你的config.py是怎么配置,无法判断你的错误,正常来说只需要改model的 feature_dim就可以了。
发送自 Windows 10 版邮件<https://go.microsoft.com/fwlink/?LinkId=550986>应用
…________________________________
发件人: cvlmm <[email protected]>
发送时间: Thursday, September 17, 2020 5:40:49 PM
收件人: yl-1993/learn-to-cluster <[email protected]>
抄送: Linsongrong <[email protected]>; Mention <[email protected]>
主题: Re: [yl-1993/learn-to-cluster] about train GCNV (#60)
@yl-1993<https://github.com/yl-1993> @Linsongrong<https://github.com/Linsongrong> ,再次请教你们
目前想训练基于512特征的GCN_V和GCN_E模型,使用的也是face_emore 5.8M的训练集,目前参考cfg_train_gcnv_ms1m来布置训练,但得到的损失一直为 losss: nan ,
@yl-1993<https://github.com/yl-1993> 请问作者和大家,我该如何配置参数训练,已得到好的结果呢
―
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub<#60 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AHFLQIC36GQVX6VJMCH56NLSGHKSDANCNFSM4PKGIX5Q>.
|
另外检查你的学习率。太大会导致模型不收敛。
发送自 Windows 10 版邮件<https://go.microsoft.com/fwlink/?LinkId=550986>应用
…________________________________
发件人: cvlmm <[email protected]>
发送时间: Thursday, September 17, 2020 5:40:49 PM
收件人: yl-1993/learn-to-cluster <[email protected]>
抄送: Linsongrong <[email protected]>; Mention <[email protected]>
主题: Re: [yl-1993/learn-to-cluster] about train GCNV (#60)
@yl-1993<https://github.com/yl-1993> @Linsongrong<https://github.com/Linsongrong> ,再次请教你们
目前想训练基于512特征的GCN_V和GCN_E模型,使用的也是face_emore 5.8M的训练集,目前参考cfg_train_gcnv_ms1m来布置训练,但得到的损失一直为 losss: nan ,
@yl-1993<https://github.com/yl-1993> 请问作者和大家,我该如何配置参数训练,已得到好的结果呢
―
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub<#60 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AHFLQIC36GQVX6VJMCH56NLSGHKSDANCNFSM4PKGIX5Q>.
|
@Linsongrong |
我不知道怎么回答你,我认为这和数据集大小关系不大。
发送自 Windows 10 版邮件<https://go.microsoft.com/fwlink/?LinkId=550986>应用
…________________________________
发件人: cvlmm <[email protected]>
发送时间: Friday, September 18, 2020 3:48:46 PM
收件人: yl-1993/learn-to-cluster <[email protected]>
抄送: Linsongrong <[email protected]>; Mention <[email protected]>
主题: Re: [yl-1993/learn-to-cluster] about train GCNV (#60)
@Linsongrong<https://github.com/Linsongrong>
嗯,感谢你的回复! 其他参数没动,改了feature_dim=512,lr=0.05,会有损失值,若默认为 lr=0.1则出现损失 loss = nan
尝试训练5.8M的数据集,以提高模型泛化能力,但完全走不动,请问你是怎么制定数据集大小的呢?,可以根据什么准确计算得到数据集限制大小呢
―
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub<#60 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AHFLQIHPUFPM77GOQC22AX3SGMGF5ANCNFSM4PKGIX5Q>.
|
或许可以考虑学习率的衰减步长,把他改小一点
发送自 Windows 10 版邮件<https://go.microsoft.com/fwlink/?LinkId=550986>应用
发件人: cvlmm<mailto:[email protected]>
发送时间: 2020年9月18日 15:48
收件人: yl-1993/learn-to-cluster<mailto:[email protected]>
抄送: Linsongrong<mailto:[email protected]>; Mention<mailto:[email protected]>
主题: Re: [yl-1993/learn-to-cluster] about train GCNV (#60)
@Linsongrong<https://github.com/Linsongrong>
嗯,感谢你的回复! 其他参数没动,改了feature_dim=512,lr=0.05,会有损失值,若默认为 lr=0.1则出现损失 loss = nan
尝试训练5.8M的数据集,以提高模型泛化能力,但完全走不动,请问你是怎么制定数据集大小的呢?,可以根据什么准确计算得到数据集限制大小呢
―
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub<#60 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AHFLQIHPUFPM77GOQC22AX3SGMGF5ANCNFSM4PKGIX5Q>.
|
@Linsongrong 嗯,正在尝试。你目前对于512GCNV网络训练取得的成绩怎样呢,引用做参考 |
1.数据集大小是有限制的,如果用gpu来训练的话,这取决与你的显存,由于是对整图进行训练,占用内存会非常大。
2.我没有对gcnv进行重新训练。
发送自 Windows 10 版邮件<https://go.microsoft.com/fwlink/?LinkId=550986>应用
发件人: cvlmm<mailto:[email protected]>
发送时间: 2020年9月18日 16:11
收件人: yl-1993/learn-to-cluster<mailto:[email protected]>
抄送: Linsongrong<mailto:[email protected]>; Mention<mailto:[email protected]>
主题: Re: [yl-1993/learn-to-cluster] about train GCNV (#60)
@Linsongrong<https://github.com/Linsongrong> 嗯,正在尝试。你目前对于512GCNV网络训练取得的成绩怎样呢,引用做参考
―
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub<#60 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AHFLQICS3XBNMAGNM3ATZZTSGMI25ANCNFSM4PKGIX5Q>.
|
@Linsongrong @yl-1993 |
@XHQC 你好, 你是对的,你对GCNV重新进行512维训练出结果了吗?效果如何?相比256维,时间花费大概增长了多少? |
@Linsongrong 256的模型我没有,故没有测试,512的模型我测试的结果是15~20分钟 50W数据,基于GCN-V,由于配置原因不稳定, |
精度变化大吗?
发送自 Windows 10 版邮件<https://go.microsoft.com/fwlink/?LinkId=550986>应用
发件人: BOD<mailto:[email protected]>
发送时间: 2020年9月29日 11:04
收件人: yl-1993/learn-to-cluster<mailto:[email protected]>
抄送: Linsongrong<mailto:[email protected]>; Mention<mailto:[email protected]>
主题: Re: [yl-1993/learn-to-cluster] about train GCNV (#60)
@Linsongrong<https://github.com/Linsongrong> 256的模型我没有,故没有测试,512的模型我测试的结果是15~20分钟 50W数据,基于GCN-V,由于配置原因不稳定,
―
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub<#60 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AHFLQIG5ODKLOWO7QRMN7LTSIFFC3ANCNFSM4PKGIX5Q>.
|
@Linsongrong 精度应该是有优势的比256模型,我的测试结果显示。 我想做5000W数据的聚类,你有没有进行过分批聚类呢,通过怎样的方式进行? |
有的,我尝试过两种方法,一种是把数据分成几部分,对每一部分数据进行全量聚类,然后再对它们的结果聚一次;第二种是先聚第一部分的结果,然后把第一部分的结果和第二部分进行聚类。在我的尝试中,第二种比第一种表现会好些,当然你可以去尝试比较两种方法,或者有更好的方法也可以分享,一起讨论。另外结果表现好坏与数据划分成反比。
发送自 Windows 10 版邮件<https://go.microsoft.com/fwlink/?LinkId=550986>应用
发件人: BOD<mailto:[email protected]>
发送时间: 2020年9月29日 11:09
收件人: yl-1993/learn-to-cluster<mailto:[email protected]>
抄送: Linsongrong<mailto:[email protected]>; Mention<mailto:[email protected]>
主题: Re: [yl-1993/learn-to-cluster] about train GCNV (#60)
@Linsongrong<https://github.com/Linsongrong> 精度应该是有优势的比256模型,我的测试结果显示。 我想做5000W数据的聚类,你有没有进行过分批聚类呢,通过怎样的方式进行?
―
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub<#60 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AHFLQIGFN6TLWHFJ3CP37B3SIFFX7ANCNFSM4PKGIX5Q>.
|
@Linsongrong 对于分批聚类再聚类你是怎样选择再聚类数据呢,是通过选择最大置信度顶点再聚类嘛,或者是? |
我选择的是类中心,当然你也可以选择置信度最大的一个或几个,我没有做过比较。或许你可以尝试一下。
发送自 Windows 10 版邮件<https://go.microsoft.com/fwlink/?LinkId=550986>应用
发件人: BOD<mailto:[email protected]>
发送时间: 2020年9月29日 11:27
收件人: yl-1993/learn-to-cluster<mailto:[email protected]>
抄送: Linsongrong<mailto:[email protected]>; Mention<mailto:[email protected]>
主题: Re: [yl-1993/learn-to-cluster] about train GCNV (#60)
@Linsongrong<https://github.com/Linsongrong> 对于分批聚类再聚类你是怎样选择再聚类数据呢,是通过选择最大置信度顶点再聚类嘛,或者是?
―
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub<#60 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AHFLQIBZZQJZL3OWAXEYSJ3SIFH3XANCNFSM4PKGIX5Q>.
|
@Linsongrong 类中心的计算方式,你是通过类均值中心值或者最近的点来选取的吧? 而置信度顶点是模型推导出的点,可以试试是否可以获取更好的类中心,最优类中心的定义我认为是正脸清晰的照片,通过计算方法得来的类中心可能会造成偏离最优类中心,这个问题貌似不好避免,由类内数据分布不均造成 |
@yl-1993 @Linsongrong @XHQC 感谢大家的真知灼见。我在使用自有数据集微调训练gcn-v的过程中发现train loss忽上忽下(可能跟我选用adam优化器有关),基本最后都会过拟合(train loss较低,但test loss超级高),而且test loss最小的模型,评估出来的指标很低啊,感觉loss失去了指导模型训练的作用。那么该怎么选择模型呢?难道每次迭代的模型都保存评估一遍?我试了下,发现在测试集指标FP能达到87%,但是其loss不是最低也不是最高,看不出规律。 |
@XHQC 你好,方便给个邮箱地址吗,我训练上有些问题想向你请教一下。 |
有512的模型么,可否共享一下,我用来测试一下聚类效果,我的特征是512维 |
@yl-1993 作者您好,我在想训练一个接受512维输入的GCN-V网络,但在配置参数时,我有些疑惑。

其中原始的nclass=1,我对这个参数配置感到疑惑,我不明白为什么是一,这不是要分类的个数吗?请您解答我的疑惑。期望您的早日答复。
The text was updated successfully, but these errors were encountered: