zyh-uaiaaaa / erasing-attention-consistency Goto Github PK

Feature visualization 那个图是怎么画出来的，可以说一下吗

Hello @zyh-uaiaaaa, thanks for sharing your work.

Can you provide AffectNet performance?

Regards,
Sungguk Cha

Thank you for your excellent work! I want to know how you train the noisy dataset? Load the MS_Celeb pretained ResNet18 model and then train noisy dataset from scratch,right?

thanks for your great work!
I can't reproduce the accuracy on both the AffectNet and FERplus datasets as mentioned in the paper. Can you provide some details of the experiment, such as hyperparameters and random number seed settings? Thank you very much.

thank you for your great work!Can you support the log file of this code ?thank you!

thanks for your great work!
can you give me the noise label of FERplus?Thank you!

Thank you very much for your work.!
I basically reproduce the performance you showed in the article on rafdb with res18 as the backbone, but for the two datasets of AffectNet and FERPlus, I can't reproduce the performance . Can you provide the parameters and random seed settings on these two datasets when the backbone is resnet18? thank you very much!

作者，你好！
我想问问数据集的问题，获取数据集的时候，给了一个basic和compound，我该用哪一个的？
还有一个 我只要用bash 运行train文件就行了？ main函数不用运行么？

Hi, expert
The proposed attention mechanism using the weights of fc layer to generate an attention map.
I would like to know setting bias is true whether to affect the attention mechanism.
Do you have any experiment about flag of bias?

您好作者！关注你们团队的研究很久了，非常感谢你们的作品，收获很多！

以resnet18为backbone、并使用数据集原始标注进行训练，我已经成功复现了89.99% acc的结果。

但是对于FERPlus，同样resnet18以作为backbone、使用相同参数的情况（batch size = 32 、learning rate = 0.0002、lambda = 5、epoch = 60、随机种子不变），使用除contempt以外的7类进行训练、测试，仍然无法复现论文中的 89.64% 的结果（我的结果仅为 85.8392%）。
我查阅了论文，似乎并没有提到训练FERPlus的具体参数、随机种子等设置，请问如何才能复现论文中的结果？

Your paper was an interesting read.

It seems the model only computes the FC layer on one view and reuses that for other views to compute the corresponding attention maps.

Note that the weights used to compute attention maps come from the FC layer

Why not computing FC separately on each view?

thanks for your great work!
I can reproduce the accuracy of the rafdb dataset when there is label noise. But I can't reproduce the SOTA results of rafdb（89.99）, can you provide some details of the experiment, such as hyperparameters and random number seed settings?

您好，请问能分享下文中用到的数据集生成的噪声标签么

'Changing backbone to ResNet-18 should first tune the learning rate in order to acquire high classification accuracy. More details can be found in the closed issues.'
您好，我浏览了下关闭的问题没有找到相关的内容。当更换backbone后，相关学习率该怎样调整呢？

Thank you for your great works!
I'm not quite clear on how the visualization map in Section 4.7 were generated(Fig. 5). Could you please share the relevant repo links or code references?
I appreciate your response to my questions.

Dear author, thank you for your great work! But all links of pretrained models in the repo which your link pointed to are dead, can you provide another link to download the pretrained model?

Hi, thanks for making the code public!
Any thoughts on providing a pre-trained code and a testing script soon?
Thanks!

Hi!
You have a memory leak during training here

It appends because
 print(correct_num) -> <add_backward>

For solving this problem, I used .detach():

loss = loss.detach().cpu()
_, predicts = torch.max(output.detach().cpu(), 1)
correct_num = torch.eq(predicts.detach().cpu(), labels.detach().cpu()).sum()

And memory stoped leak.

I attach a memory profile file.

你好，作者，能提供一下基础resnet50训练到88.75的代码么？还有mobileNet训练到83.31的代码么？在RAF-DB数据集上的

Did MobileNet in Table 2 use a pre-trained model? Can you provide the download link for this pre-trained model?