question about the version of pytorch

Question

question about the version of pytorch

zhiqiangzhongddu opened this issue 5 years ago · comments

Hi,

Thanks for your kind sharing.
I want to reproduce your results on my device, but my training results are too bad,
The evaluating HR and NDCG of some epochs are 0.0000, do you have any idea about this case? And could you tell me the Pytorch version that you use in your side?

Thanks.

Yihong Chen · Answer 1 · Wed Mar 20 2019 10:12:52 GMT+0800 (China Standard Time)

Could you share your training curve? I tested the repo using pytorch 0.2 but I guess it should work under 0.3 and 0.4 as well.

GabbyHE · Answer 2 · Wed Mar 20 2019 14:05:30 GMT+0800 (China Standard Time)

Hi,

Thanks for your kind sharing.
I want to reproduce your results on my device, but my training results are too bad,
The evaluating HR and NDCG of some epochs are 0.0000, do you have any idea about this case? And could you tell me the Pytorch version that you use in your side?

Thanks.

@LaceyChen17 @zhiqiangzhongddu , I got the same issue, see the curves on dataset ml-1m,

Zhiqiang ZHONG · Answer 3 · Wed Mar 20 2019 16:51:13 GMT+0800 (China Standard Time)

Could you share your training curve? I tested the repo using pytorch 0.2 but I guess it should work under 0.3 and 0.4 as well.

Before, I used Pytorch 1.0.1, the latest version.
I tried Pytorch 0.4.0 with Cuda 80&90, but they don't work.
Error:
RuntimeError: input and target shapes do not match: input [1024 x 1], target [1024].

And I'm not success to install Pytorch 0.2 and 0.3 on my computer. It seems Conda doesn't support them anymore.

Yihong Chen · Answer 4 · Wed Mar 20 2019 16:54:01 GMT+0800 (China Standard Time)

I reproduced the bug in pytorch 1.0 as well. I am working on it now. Yihong

…

________________________________ 发件人: Zhiqiang ZHONG <notifications@github.com> 发送时间: 星期三, 三月 20, 2019 4:51 下午收件人: LaceyChen17/neural-collaborative-filtering 抄送: Yihong Chen; Mention 主题: Re: [LaceyChen17/neural-collaborative-filtering] question about the version of pytorch (#5) Could you share your training curve? I tested the repo using pytorch 0.2 but I guess it should work under 0.3 and 0.4 as well. Before, I used Pytorch 1.0.1, the latest version. I tried Pytorch 0.4.0 with Cuda 80&90, but they don't work. Error: RuntimeError: input and target shapes do not match: input [1024 x 1], target [1024]. And I'm not success to install Pytorch 0.2 and 0.3 on my computer. It seems Conda doesn't support them anymore. ― You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub<#5 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/ALNdqxT9W3QWAZywswnuowYjVAIyNjN8ks5vYfaCgaJpZM4b8752>.

Zhiqiang ZHONG · Answer 5 · Wed Mar 20 2019 17:24:47 GMT+0800 (China Standard Time)

news: the same bug exists with Pytorch-0.3.1, Cuda=90&80&100 in my computer.

Yihong Chen · Answer 6 · Thu Mar 21 2019 19:46:36 GMT+0800 (China Standard Time)

@zhiqiangzhongddu @GabbyHE

The bug was mainly caused by MSELoss.

The repo works well under the latest pytorch(1.0.1post2) now.

Thank you for reporting the bug. Also, pay attention to l2_regularization parameter.

Strong l2_regularization might lead to model underfitting and thus HR 0 & NDCG 0.

GabbyHE · Answer 7 · Thu Mar 21 2019 20:10:34 GMT+0800 (China Standard Time)

@zhiqiangzhongddu @GabbyHE

The bug was mainly caused by MSELoss.

The repo works well under the latest pytorch(1.0.1post2) now.

Thank you for reporting the bug. Also, pay attention to l2_regularization parameter.

Strong l2_regularization might lead to model underfitting and thus HR 0 & NDCG 0.

Thanks so much for fixing this bug.

Zhiqiang ZHONG · Answer 8 · Mon Mar 25 2019 17:35:04 GMT+0800 (China Standard Time)

@zhiqiangzhongddu @GabbyHE

The bug was mainly caused by MSELoss.

The repo works well under the latest pytorch(1.0.1post2) now.

Thank you for reporting the bug. Also, pay attention to l2_regularization parameter.

Strong l2_regularization might lead to model underfitting and thus HR 0 & NDCG 0.

Thanks.