kaiwang960112 / challenge-condition-fer-dataset Goto Github PK

浏览了您的代码与论文，我想请问一下您的ferplus数据集部分的训练问题。
首先关于ferplus三万多张的48x48的照片您全部用dlib裁剪后对齐了嘛，然后训练数据全部借助caffecrop类resize到224224大小吗？
其次是关于您给出的一些训练用文件，如dlib_ferplus_train_center_crop_range_label.txt，它们是在全部在原始数据集上三万多张照片上对齐生成的吗，筛选掉了哪些照片不用于训练呢？
最后是想请教给出的遮挡和多角度的照片是怎样筛选出来的呢，它的编号意味着什么呢，为什么都是003***.png的格式呢，更小编号的照片也存在遮挡或多角度的问题。
还劳请您在工作之余能帮助解疑答惑，感谢您！

Hi,

Could you please also share the image list of the occlusion RAF-DB dataset? Thank you very much!

0/fer0000001 6
0/fer0000002 6
0/fer0000004 6
0/fer0000005 6

可以看到，你的txt里将1 2 4 5 分成了一组，但实际上1 2 4 5 这4张图片在官方csv中对应分类是0 2 6 2。请问你是对数据集的标签进行了什么修改吗？

您好，我读了您的论文，并尝试运行您提供的代码，我有两个问题（我用的是FER的code）

1. 是必须所有种类的图片格式差别不能太大吗？我发现差别超过batch的话，就会报错
2. train_aatention_loss_rank.py 240行附近为什么target赋值为target_first?

您好，感谢您在百忙之中查看我的issue。我在load您开源的模型时了一些问题，load模型的代码和出现的错误信息见下：
代码：
`all_data=torch.load('../../ijba_res18_naive.pth')

model=resnet18(end2end=False)

model.load_state_dict(all_data['state_dict'])

print('model',model)`

错误信息：
Missing key(s) in state_dict: "conv1.weight", "bn1.weight", "bn1.bias",......
Unexpected key(s) in state_dict: "module.conv1.weight", "module.bn1.weight", "module.bn1.bias", ....

希望得到您的解答，谢谢。

作者你好,我在test_rank_loss里发现要import这个文件,能放出来一份val_part_attention 这个文件吗?多谢

作者您好！
我在icip2020上看到了您的论文，我现在想尝试投一下icip2021，请问您当时投稿时选择的session topic是啥呀？我没看到直接的表情识别相关的

I want to ask about the pretrained model. Your model define in this repo include alpha and beta layer, but it's not found on pretrained model, how can i load weight for that module (i have to load model without alpha and beta)? Or maybe I misunderstand some where?
(i'm using ijba_res18_naive.pth.tar pretrained model)

Following files are missing in FERplus_dir.

• attention_rank_loss.py
• test_rank_loss_attention.py
• val_part_attention_sample.py

I will also appreciate to include the code to crop the images.

作者您好！
我尝试运行您提供的代码，但是我有些问题关于您提到的局部crop，在FERplus路径下的part_attetion_sample.py的173行，这里是不是没有实现crop的操作，我在这里设置了断点并输出图片，图片和原图是一样的，请问是我对程序理解错误了吗？

作者您好，最近看了您这篇关于RAN的文章，想要试着跑一下你提供代码，但是在运行ferplus_dir/test_rank_loss_attention时遇见了一时无法理解的问题，特地向你请教下。
1.train_attention_rank_loss.py中每个epoch都会输出一个TEST，请问之后的test_rank_loss_attention.py的功能又有什么不同呢？
2.能否提供一下，test_rank_loss_attention.py中61行附近的./data/resnet18/checkpoint_40.pth.tar包或者获取方式
3.能否提供一下，test_rank_loss_attention.py中74行附近txt_path = '/media/sdc/kwang/ferplus/pose_test/test_txt/' ，里面的pose_test中的内容
4.并没有发现这几个.py文件中使用了pose_30_ferplus_list.txt , pose_45_ferplus_list.txt, jianfei_occlusion_list.txt
非常感谢，再加实在抱歉，新人跑模型遇见很多问题，加上对作者的工程目录不清晰，遇见了很多让大佬无法理解的问题，轻喷轻喷轻喷

作者你好，最近看了你的论文，想复现你的实验，但没有预训练模型和数据集，你的预训练模型和数据集哪里可以下载呢？没有看到你说的百度网盘链接呢？

Hi, thank you for your excellent work!

I am trying to reproduce the training results on the full FER_Plus dataset by using the code (train_attention_rank_loss.py) and pretrained resnet18 model (ijba_res18_native.pth.tar) you provided. I also use the fixed cropping strategy as in your paper (full image + 5 regions). Besides the learning rate, all of the training settings are unchanged.

However, I could not achieve the accuracy mentioned in the paper. Most of the time, the validation and testing accuracy are only around 85% and 83%, respectively. The model can fit pretty well on the training set, though.

Do you have any special preparation for setting up the training or any configuration of the hyperparameters? Your suggestion would be highly appreciated!

作者您好，您这篇文章非常精彩。我是一个初学者，对于将裁剪的图片送入resnet，其输出经过一个fc和sigmoid就得到该区域的注意力权重比较困惑。为什么这个值就是该区域的attention weight？另外本文中的self-attention和transformer中的self-attention是一个东西吗?水平有限，还望赐教。

运行AffectNet_dir的simple——sample的train_attention_part_face.py的时候出现这个错误，是为什么啊？这个文件是在哪里啊

FileNotFoundError: [WinError 3] 系统找不到指定的路径。: '/media/sdd/kwang/affectnet/fly_part_affectnet/train/1/737db2483489148d783ef278f43f486c0a97e140fc4b6b61b84363ca'

Hi, what's the license for the code/models? thanks :)

作者，您好，最近您的文章，想要试着跑一下你提供代码，但是在运行ferplus_dir / test_rank_loss_attention时发现缺少val_part_attention，请问这个文件可以在哪里获取呢

您好，我对您的代码特别感兴趣，最近临近毕业，想参考您的代码，您能否分享一份详细的关于RAF-DB数据集的代码，谢谢！

询问在区域自注意力确定权重时，论文中的q0 ql如何得出，体现在代码的哪个部分？谢谢

It is a great work!

In your paper, Setting the radius as 0.4 of the side of image for landmark-based cropping.I just find the max and min of xy coordinates, then add pad.But I don't know how to crop using a radius.Can you give a example? Thanks a lot!

感谢您的工作和开源代码~

运行的时候发现缺少了crop部分的代码，查看了相关的issue，发现您提到在dataset文件夹中，但代码库好像没看到dataset文件夹，请问下您是否方便分享下crop部分的代码，不方便也麻烦告知下，感谢~

I read your paper with great interest.
thank you.

And I want to use the dataset you deployed, but I don't know how to use Affect-net with occlusion.

The affect-net I downloaded is composed of affectnet/val/emotion number/images number, and your occluded affecnet txt file is affectnet/val/emotion number/bc9ad400db82eca1882b11c9603348fb3e92f5b0d67391ac22a1bc79/. The last part(images name) seems to be different from the format of the published dataset.

How should I convert files name and use it?

It would be of great help if you could tell me how to use it.

Hi,

Thank you very much for sharing the awesome work. Recently, I am reading your paper, and have a question: for all the experiment results, do you use the pre-trained face recognition network as initialization? Thanks!

Occlusion list里面每个图片名后有两个label，第一列的是表情label，第二列的是遮挡类型标签。

作者您好，首先膜拜一波大神，祝您CVPR顺利搞定。
我最近在研究解决人脸表情识别中的遮挡和姿态问题，很需要您论文中提到的6个数据集Occlusion-FERPlus, Pose-FERPlus, Occlusion-AffectNet, Pose-AffectNet, Occlusion-RAF-DB, 和 Pose-RAF-DB。请问您是否可以分享一下呢，非常感谢~！

Thanks for answering my last question. I really like your work. Could you also share the best trained model on Google Drive/OneDrive? I have been trying for hours and I still couldn't download it from baidu for some reason.

Cannot open the state-of-the-art model. The file is corrupted for some reason. Would you please by any chance fix it? Thanks!

As far as I know FER+ provides images not videos. That being said, I don't quite understand why each face is represented as video in part_attention_sample.py. I'm gussing, If fixed position cropping is used, does this mean video_path contains 6 images with different part + original image? Or how do you get the videos and extract frames?

img_path_first = video_path+'/'+img_lists[0]
img_path_second = video_path + '/'+img_lists[1]
img_path_third = video_path + '/'+img_lists[2]
img_path_forth = video_path + '/'+img_lists[3]
img_path_fifth = video_path + '/' + img_lists[4]
img_path_sixth = video_path + '/' + img_lists[5]

切完后的人脸经过ToTensor()中的除以255.的归一化就输入网络了么

作者好，我完成了ferplus数据集的模型训练，现在想进行RAF数据集的训练，然后再复现自愈网络SCN，但是我并没有找到RAF数据集的关键代码，想看你是否方便将RAF的源码分享给我呢？我的邮箱是[email protected]，如果有打扰到你，希望得到你的谅解，期待收到你的回复，不胜感激。

还有个文件val（pos>30）里面文件只有1247个标签，论文写有1248个

I have sent two e-mails to two e-mail address published at http://www.whdeng.cn/raf/model1.html. But no respond receive. I need your help. Thank you a lot.

您好，感谢您开源代码，但是在复现过程中，我发现您并没有提供在test_rank_loss_attention.py 中的 "ferplus/pose_test/test_txt/"文件夹，此外，您在另外一个issue中说已经提供了，是您忘记上传了吗？

此外，您当前pytorch中部分代码已经在2020年以来的pytorch版本中弃用了。如果您能提供上面那个文件夹使我完成复现的话，我同样也希望可以为这个repo做出贡献，也希望commit一个新版本的代码，外加简单的image crop。期待您的回复~

This is the last layer.
module.fc.weight', (87020, 256)
module.fc.bias', (87020,)
why the num_classes is 87020 ? And I can not load it properly.
Thank you for watching.

Hello.

THanks for your code. I've tried to use your uploaded model "ijba_res18_naive.pth.tar" to infer using script FERplus_dir/test_rank_loss_attention.py, but i've got error on the line

model.load_state_dict(checkpoint['state_dict'])

Here is the error

RuntimeError: Error(s) in loading state_dict for DataParallel:
	Missing key(s) in state_dict: "module.alpha.0.weight", "module.alpha.0.bias", "module.beta.0.weight", "module.beta.0.bias". 
	Unexpected key(s) in state_dict: "module.feature.weight", "module.feature.bias". 
	size mismatch for module.fc.weight: copying a param with shape torch.Size([87020, 256]) from checkpoint, the shape in current model is torch.Size([8, 512]).
	size mismatch for module.fc.bias: copying a param with shape torch.Size([87020]) from checkpoint, the shape in current model is torch.Size([8]).

Probably i'm doing something wrong and this file "ijba_res18_naive.pth.tar" should be used somewhere else. Will be glad to hear some advises.

在TABLE VII中你将你们的算法和[44], [58]对比，据我所知[44]中report的是8个表情的平均准确率，[58]中为了和raf-db融合去除了表情contempt，report的是7个基本表情的平均准确率。另外[44]和[58]均未在人脸数据集上pretrian，这对表情结果识别影响不小。想咨询两个问题：
（1）你的59.5是8个表情的平均准确率还是7个表情的平均准确率？即有无包含contempt表情。
（2）你的59.5是否是在pretrain的人脸识别模型上训练而来？
多谢！

Hello there, the dataset I got after request has 15339 total images in basic but the web page for RAFDB says it is around 30,000 images. Could you please tell me number of images you trained your model on? I could not find any useful information anywhere on web.

请问你们的数据集有用来举办什么比赛吗？或者刷榜工作？

你好，请问一下你这个模型分了几类表情呢，好像不止7类？

您好，请问FERplus_dir目录下的val_part_attention.py是要自己到别的地方下载还是说忘放进去呢？谢谢

作者您好！关注了您在TIP和CVPR上发表的论文RAN与SCN，现在准备根据您提供的github 开源代码复现RAN，据我所知，在RAN论文中提到在FERPlus上最高的点数是89.16，而且是以VGG16为backbone实现的，但是在开源的代码中没有基于VGG16的网络实现，请问您是否可以开源这一块的代码？此外，我想请问在FERPlus，您是否有尝试resnet50、VGG19等网络层次更深的网络作为backbone？它们有提高FER的准确度吗？期待您的回复，谢谢！

我想问一下，本程序中数据集应该整理为什么形式呢，是每个文件夹里放一种表情吗，然后为每个原始图片创建一个文件夹，里边放剪好的图片？

Hi,
Can you please clarify:
Is ijba_res18_naive.pth model finetuned on FER datasets or just pre-trained only on FR task only? I wish to used pretrained model for a new model. Can you please point me to where to get resnet-18 pretrained on MSceleb?
Thanks

您好，请问您可以给一下checkpoint_[1, 1]_33_0.002_89.162.pth.tar 的网络结构代码吗？想加载这个vggface的预训练模型，但是不确定网络结构如何？里面好像有drop7.weight和drop7.bias，不确定是什么结构？

Hi there,
Thanks for your excellent work. I can't find any code for cropping the images. Although there is a class namely "CaffeCrop" in several scripts, it only augments the data not cropping them. As mentioned in the paper, there should be three ways for cropping (random, consistent, and landmark-based) but I can't find them in the code provided. Can you please share it? Or if it is in the repository, please guide me on which one is?
Looking for a response :)
Thank you