zyh-uaiaaaa / relative-uncertainty-learning Goto Github PK

Hello,Thank you so much for your excellent work,but I can't find it download ms-celeb pretrained model for weight initialization，because the pre-trained model provided by the author is not linked to Google Drive:https://drive.google.com/file/d/1H421M8mosIVt8KsEWQ1UuYMkQS8X1prf/view?usp=sharing

I want to retrain the model, so I would like to ask you to re-upload the pre-trained model to Baidu NetDisk or Google Drive(resnet18_msceleb.pth),thank you very much.

Thanks for your excellent work! I have a question about the synthetic noisy experiment.
As described in your article 【We run all experiments three times and compute the mean and standard variance of the results】
I wonder if you choose the best result of three experiments and then compute the mean and standard variance, or some other way?

图二看起来是input层面上的mixup，而原文中使用的是feature mixup，请问怎么理解

Dear authors,

thank you for sharing the codes of your nice work! However, when we tried the code, we found that the results are not very consistent on different platforms. Specifically, we made three trials (without any modification to the source code):
[1] On a single GTX 1080Ti, the final accuracy is 88.85;
[2] On a single RTX 3090, the final accuracy is 88.33;
[3] On a single A40, the final accuracy is 88.62.

We are curious why this phenomenon happens. Also, we wonder if the authors could provide some suggestions on reproducing the results. Thank you!

Does the pre-training model mentioned in README need to be trained? Or it can be used to test directly？

您好！非常感谢提供这么好的不确定估计学习工作。我想请教您，RUL方法可以用于多标签分类么？如果想用类似的不确定性估计方式解决多标签分类问题，有什么建议么？另外，我看到文中采用不同label 图像的feature进行mixup，那如果是随机的img的feature 进行mixup操作，那么性能如何呢？

期待您的答复。

祝好！

Thanks for your great work. But i have a question, can you help me how to get the uncertainty values of input image ?

Hi, thanks for sharing the source codes.
Can we also have the code for evaluation of the pre-trained model? I just can see the main.py file which is for training the model.
Would be appreciated if you can share the test/evaluation code as well.
Thanks.

As a emotion researcher, i have the similar thought on the DUL on CVPR2020. And I found your paper on list of NIPS,which makes me excited.

However, I could not find your paper on arxiv... Can you tell me where I can find it?

Also, I am a BYR, and a FER&FR researcher too. :)

作者您好，感谢您分享的论文代码，我在尝试将FER2013用于您的代码时，怎么样都达不到目标准确率，所以我想请问一下，这个数据集是需要做什么预处理吗？如果可以的话，希望您可以指导一下

Thanks for your great work.

I have a question and need your help.
How to calculate the mean +/- standard deviation (like 80.43±0.72) in "Table 1: Test accuracy (%) on RAF-DB, FER2013 and AffectNet with synthetic noisy labels."?
Is it to record the maximum accuracy in each experiment, and then calculate the mean and standard deviation of multiple experiments?

Hope to hear from you soon. TQ

Could you share the code, seeds that generate random noisy labels (10%, 20%, 30%) as mentioned in the paper?