Closed mobulan closed 1 year ago
Sure, these days I am preparing for the paper works about PhD thesis, so I guess that I would submit codes before 15 June. Thank you for focusing our works.
Thank you very much for your work, also hope to release the code of this project as soon as possible.
Thank you very much for your work, also hope to release the code of this project as soon as possible.
So sorry for my late submission. I just finish my vacation with my mum. I will do this this week.
Thank you very much for your work, also hope to release the code of this project as soon as possible.
So sorry for my late submission. I just finish my vacation with my mum. I will do this this week.
Thank you very much!
I checked the code. The codes include all things you mentioned, although two of them are not expected with the unbalance setting and the version without OT (I will fix this week). So what's your question? Maybe you misunderstand the code?
Mobulan @.***> 于2023年7月12日周三 19:28写道:
The currently published code is not complete, especially the codes of the label propagation algorithm, prototype construction, training process, and related parameter settings are missing. This is very important for us to further understand your method. Can you provide as much as possible, thank you very much!
— Reply to this email directly, view it on GitHub https://github.com/allenhaozhu/protoLP/issues/1#issuecomment-1632166164, or unsubscribe https://github.com/notifications/unsubscribe-auth/AMPJZOEWI3AGFH3NDQIMHN3XPZU3HANCNFSM6AAAAAAY56BJ54 . You are receiving this because you commented.Message ID: @.***>
I'm sorry, I misunderstood the code and the content of the text, this code does contain all the content, I have deleted the related issues on github, thank you very much for your reply. In addition, under my current pytorch 2.0 settings, the accuracy of the reproduction will be about 0.2% lower than that marked in the text, which is the same as iLPC. Do I need to make any additional settings?
从 Windows 版邮件发送
发件人: allenhaozhu 发送时间: 2023年7月13日 20:43 收件人: allenhaozhu/protoLP 抄送: Mobulan; Author 主题: Re: [allenhaozhu/protoLP] Looking forward to your code (Issue #1)
I checked the code. The codes include all things you mentioned, although two of them are not expected with the unbalance setting and the version without OT (I will fix this week). So what's your question? Maybe you misunderstand the code?
Mobulan @.***> 于2023年7月12日周三 19:28写道:
The currently published code is not complete, especially the codes of the label propagation algorithm, prototype construction, training process, and related parameter settings are missing. This is very important for us to further understand your method. Can you provide as much as possible, thank you very much!
— Reply to this email directly, view it on GitHub https://github.com/allenhaozhu/protoLP/issues/1#issuecomment-1632166164, or unsubscribe https://github.com/notifications/unsubscribe-auth/AMPJZOEWI3AGFH3NDQIMHN3XPZU3HANCNFSM6AAAAAAY56BJ54 . You are receiving this because you commented.Message ID: @.***>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>
sorry, I am busy finding a job these days. About the performance with Sinkhorn distance, we have mentioned in https://zhuanlan.zhihu.com/p/633999016. It is a trick for all methods using Sinkhorn distance, that is why we propose this method. In unbalanced setting, all these methods will be crashed. The 0.2 ACC gap is from some dataset-specific tricks such as renormalization anchors with L2 normalization. Thanks for your attention.
Mobulan @.***> 于2023年7月14日周五 00:50写道:
I'm sorry, I misunderstood the code and the content of the text, this code does contain all the content, I have deleted the related issues on github, thank you very much for your reply. In addition, under my current pytorch 2.0 settings, the accuracy of the reproduction will be about 0.2% lower than that marked in the text, which is the same as iLPC. Do I need to make any additional settings?
从 Windows 版邮件发送
发件人: allenhaozhu 发送时间: 2023年7月13日 20:43 收件人: allenhaozhu/protoLP 抄送: Mobulan; Author 主题: Re: [allenhaozhu/protoLP] Looking forward to your code (Issue #1)
I checked the code. The codes include all things you mentioned, although two of them are not expected with the unbalance setting and the version without OT (I will fix this week). So what's your question? Maybe you misunderstand the code?
Mobulan @.***> 于2023年7月12日周三 19:28写道:
The currently published code is not complete, especially the codes of the label propagation algorithm, prototype construction, training process, and related parameter settings are missing. This is very important for us to further understand your method. Can you provide as much as possible, thank you very much!
— Reply to this email directly, view it on GitHub https://github.com/allenhaozhu/protoLP/issues/1#issuecomment-1632166164,
or unsubscribe < https://github.com/notifications/unsubscribe-auth/AMPJZOEWI3AGFH3NDQIMHN3XPZU3HANCNFSM6AAAAAAY56BJ54>
. You are receiving this because you commented.Message ID: @.***>
— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>
— Reply to this email directly, view it on GitHub https://github.com/allenhaozhu/protoLP/issues/1#issuecomment-1634384361, or unsubscribe https://github.com/notifications/unsubscribe-auth/AMPJZOG5UKRBB67HAZXU3XLXQADK5ANCNFSM6AAAAAAY56BJ54 . You are receiving this because you commented.Message ID: @.***>
maybe there is a small mistake for the file name? I can re-implement your result with the file “without _full”, maybe this version use the sinkhorn distance发自莫布兰的 iPhone在 2023年7月21日,08:51,allenhaozhu @.***> 写道: sorry, I am busy finding a job these days. About the performance with
Sinkhorn distance, we have mentioned in
https://zhuanlan.zhihu.com/p/633999016. It is a trick for all methods using
Sinkhorn distance, that is why we propose this method. In unbalanced
setting, all these methods will be crashed. The 0.2 ACC gap is from some
dataset-specific tricks such as renormalization anchors with L2
normalization. Thanks for your attention.
Mobulan @.***> 于2023年7月14日周五 00:50写道:
I'm sorry, I misunderstood the code and the content of the text, this code
does contain all the content, I have deleted the related issues on github,
thank you very much for your reply.
In addition, under my current pytorch 2.0 settings, the accuracy of the
reproduction will be about 0.2% lower than that marked in the text, which
is the same as iLPC. Do I need to make any additional settings?
从 Windows 版邮件发送
发件人: allenhaozhu
发送时间: 2023年7月13日 20:43
收件人: allenhaozhu/protoLP
抄送: Mobulan; Author
主题: Re: [allenhaozhu/protoLP] Looking forward to your code (Issue #1)
I checked the code. The codes include all things you mentioned, although
two of them are not expected with the unbalance setting and the version
without OT (I will fix this week). So what's your question? Maybe you
misunderstand the code?
Mobulan @.***> 于2023年7月12日周三 19:28写道:
The currently published code is not complete, especially the codes of
the
label propagation algorithm, prototype construction, training process,
and
related parameter settings are missing. This is very important for us to
further understand your method. Can you provide as much as possible,
thank
you very much!
—
Reply to this email directly, view it on GitHub
https://github.com/allenhaozhu/protoLP/issues/1#issuecomment-1632166164,
or unsubscribe
<
https://github.com/notifications/unsubscribe-auth/AMPJZOEWI3AGFH3NDQIMHN3XPZU3HANCNFSM6AAAAAAY56BJ54>
.
You are receiving this because you commented.Message ID:
@.***>
—
Reply to this email directly, view it on GitHub, or unsubscribe.
You are receiving this because you authored the thread.Message ID:
@.***>
—
Reply to this email directly, view it on GitHub
https://github.com/allenhaozhu/protoLP/issues/1#issuecomment-1634384361,
or unsubscribe
https://github.com/notifications/unsubscribe-auth/AMPJZOG5UKRBB67HAZXU3XLXQADK5ANCNFSM6AAAAAAY56BJ54
.
You are receiving this because you commented.Message ID:
@.***>
—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you modified the open/close state.Message ID: @.***>
in addition, there is an error when use the resnet-12 backbone in "without _full" file. i comment the line 293 and the problem can be resolved (ndatas[:,] = torch.pow(ndatas[:,]+1e-6, beta)
) and the results can also be reimplemented.
Thank you very much for your work, hope to release the code of this project as soon as possible.