Why "iface_preds " contain NAN when training dmasif_

[BingzeWu]

It seems that the training is not stable.
I follow the "benchmark_scripts" to retrain the dMaSIF_site_3layer_9A. But when calculate the roc-auc, it raised "Problem with computing roc-auc" and I found that the "iface_preds" contain NAN.
Does anyone have similar problem?

[YAndrewL]

Same problem, did you solve this? @BingzeWu

[YAndrewL]

What do you mean by mini-batch? I've trained this with a batch size of 64, but the model only considers single-batch training, and NaN values still appear after several steps. Bingze Wu ***@***.***> 于2023年11月10日周五 14:14写道：

…

No, I found the problem may come from the data preprocess step. When I trained the model on a mini batch, the training was successful. So I found in dMasif convolution step, there is some problem for the “nuv” data. But I don’t how to fix the bug. ***@***.*** 发件人: Yufan Andrew Liu ***@***.***> 日期: 星期三, 2023年11月8日 23:59 收件人: FreyrS/dMaSIF ***@***.***> 抄送: Bingze WU 吴秉泽 ***@***.***>, Mention ***@***.***> 主题: Re: [FreyrS/dMaSIF] Why "iface_preds " contain NAN when training dmasif_ (Issue #48) 你通常不会收到来自 ***@***.*** 的电子邮件。了解这一点为什么很重要< https://aka.ms/LearnAboutSenderIdentification> Same problem, did you solve this? @BingzeWu<https://github.com/BingzeWu> ― Reply to this email directly, view it on GitHub< #48 (comment)>, or unsubscribe< https://github.com/notifications/unsubscribe-auth/A2G24DEQKR5UEZBMBUARQ7TYDOT3LAVCNFSM6AAAAAA6LQCAYSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQMBSGE4DKNRTHE>. You are receiving this because you were mentioned.Message ID: ***@***.***> — Reply to this email directly, view it on GitHub <#48 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AOTEIWBOWVUVXJ7QA7TJFV3YDYEBBAVCNFSM6AAAAAA6LQCAYSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQMBVGUZTGMJQGQ> . You are receiving this because you commented.Message ID: ***@***.***>

-- Yufan Liu, Ph.D. student in computer science, Computational Bioscience Research Center (CBRC), King Abdullah University of Science and Technology (KAUST) yandrewl.github.io

[BingzeWu]

Sorry, I mean I trained the model on a sub-dataset(randomly chosen, about 300 data points). And when training on the complete dataset, NaN values still appeared. If you use the trained model to compute the problem date point(where roc-auc problem was raised), you will find in specific geometric convolution layers the Nan value appeared. The internal computation seems to give the wrong value, but I don’t know how to fix it. The convolution relies on different geometric operations, which I am unfamiliar with. 发件人: Yufan Andrew Liu ***@***.***> 日期: 星期一, 2023年11月13日 19:32 收件人: FreyrS/dMaSIF ***@***.***> 抄送: Bingze WU 吴秉泽 ***@***.***>, Mention ***@***.***> 主题: Re: [FreyrS/dMaSIF] Why "iface_preds " contain NAN when training dmasif_ (Issue #48) 你通常不会收到来自 ***@***.*** 的电子邮件。了解这一点为什么很重要<https://aka.ms/LearnAboutSenderIdentification> What do you mean by mini-batch? I've trained this with a batch size of 64, but the model only considers single-batch training, and NaN values still appear after several steps. Bingze Wu ***@***.***> 于2023年11月10日周五 14:14写道：

…

No, I found the problem may come from the data preprocess step. When I trained the model on a mini batch, the training was successful. So I found in dMasif convolution step, there is some problem for the “nuv” data. But I don’t how to fix the bug. ***@***.*** 发件人: Yufan Andrew Liu ***@***.***> 日期: 星期三, 2023年11月8日 23:59 收件人: FreyrS/dMaSIF ***@***.***> 抄送: Bingze WU 吴秉泽 ***@***.***>, Mention ***@***.***> 主题: Re: [FreyrS/dMaSIF] Why "iface_preds " contain NAN when training dmasif_ (Issue #48) 你通常不会收到来自 ***@***.*** 的电子邮件。了解这一点为什么很重要< https://aka.ms/LearnAboutSenderIdentification> Same problem, did you solve this? @BingzeWu<https://github.com/BingzeWu> D Reply to this email directly, view it on GitHub< #48 (comment)>, or unsubscribe< https://github.com/notifications/unsubscribe-auth/A2G24DEQKR5UEZBMBUARQ7TYDOT3LAVCNFSM6AAAAAA6LQCAYSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQMBSGE4DKNRTHE>. You are receiving this because you were mentioned.Message ID: ***@***.***> ― Reply to this email directly, view it on GitHub <#48 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AOTEIWBOWVUVXJ7QA7TJFV3YDYEBBAVCNFSM6AAAAAA6LQCAYSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQMBVGUZTGMJQGQ> . You are receiving this because you commented.Message ID: ***@***.***>

-- Yufan Liu, Ph.D. student in computer science, Computational Bioscience Research Center (CBRC), King Abdullah University of Science and Technology (KAUST) yandrewl.github.io ― Reply to this email directly, view it on GitHub<#48 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/A2G24DFYQ67MUK7PNSRN7CDYEIANJAVCNFSM6AAAAAA6LQCAYSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQMBXHE4DIOJWHA>. You are receiving this because you were mentioned.Message ID: ***@***.***>

[Xinheng-He]

I also found that it has issues with input data, batch size and its hyperparameters. Struggled for 1 week for running without nan but fails. Maybe such network only works for their own PPI data which passed through prescision regulation. Given up for understanding and debug it.

[orange2350]

@YAndrewL Same problem, did you solve this?

[YAndrewL]

Hi Zhiyi, not yet, but you may find the NaN in the input feature part, and mask then with average or some constant to start the training, unfortunately, I did not get the training results described in the paper. Chen Zhiyi ***@***.***> 于2023年12月6日周三 05:44写道：

…

@YAndrewL <https://github.com/YAndrewL> Same problem, did you solve this? — Reply to this email directly, view it on GitHub <#48 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AOTEIWAQUAMMBJH7ACXOSHLYH7LYFAVCNFSM6AAAAAA6LQCAYSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNBRHE4TKNZUGE> . You are receiving this because you were mentioned.Message ID: ***@***.***>

-- Yufan Liu, Ph.D. student in computer science, Computational Bioscience Research Center (CBRC), King Abdullah University of Science and Technology (KAUST) yandrewl.github.io

[orange2350]

Hi Zhiyi, not yet, but you may find the NaN in the input feature part, and mask then with average or some constant to start the training, unfortunately, I did not get the training results described in the paper. Chen Zhiyi @.> 于2023年12月6日周三 05:44写道：
…
@YAndrewL https://github.com/YAndrewL Same problem, did you solve this? — Reply to this email directly, view it on GitHub <#48 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOTEIWAQUAMMBJH7ACXOSHLYH7LYFAVCNFSM6AAAAAA6LQCAYSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQNBRHE4TKNZUGE . You are receiving this because you were mentioned.Message ID: @.>
-- Yufan Liu, Ph.D. student in computer science, Computational Bioscience Research Center (CBRC), King Abdullah University of Science and Technology (KAUST) yandrewl.github.io

The results I get from reproducing the dMASIF is not the same as the paper used to evaluate it either.thanks