复现精度远达不到论文数据，作者为何不公布训练代码？是否使用了测试数据训练模型？

[jessicaaaio1993]

作者始终不肯公布关键训练代码，并且对相关问题视而不见。我们只能自己复现训练代码，结果远远达不到论文精度（如下表所示）。请作者正面回应问题并解释为何只公开测试代码。作者是否使用了部分测试数据来训练模型？

WFLW数据集	NME $\downarrow$	FR 0.1 $\downarrow$	AUC 0.1 $\uparrow$
作者原文（HRNetW18C-lite）	4.14	2.76	0.595
我们复现训练代码	4.67	4.09	0.581

300W数据集NME $\downarrow$	Common	Challenging	Fullset
作者原文（HRNetW18C-lite）	2.75	4.90	3.17
我们复现训练代码	2.95	5.27	3.50

[Jiahao-UTS]

Thanks for your interest. We only used the training set. You can find the training log at https://drive.google.com/file/d/1ZWUGq_11RwWOVSvojJZ8ttNJkgyNSaaj/view?usp=sharing. If you have problem, I think you should discuss the problem with us firstly, rather than register a new account and publish such aggressive words. Anyway, I never refuse publish the training code. Because of some issues, the training code is planned to be released in 2023. If you cannot reproduce the results, I think you should check your code firstly. Thank you!

[jessicaaaio1993]

Hi, author! Nice to see your first reply to the issue related to the missing training code. Unfortunately, you're still trying to prevaricate the problem. Many people mentioned the problem with the training code but you never replied to them. Can you respond to the question directly? It has been nearly a year since the paper was accepted, what caused you to procrastinate until now and not release the training code? You said that it would be released in 2023, which is very strange. Can you give us a specific deadline? Once the training code is published and can achieve the performance in the paper, I will apologize to you and retract all my comments. Or I will consider this as academic misconduct behavior and report this to the conference committee.