hcplab-sysu / cd-fer-benchmark Goto Github PK

你好，我下载您指定的数据集，发现CD-FER-Benchmark/Dataset/CK+/CK+_Emotion/Train/CK+_Train_crop/Angry文件夹下面是空的，没有数据，麻烦解答。谢谢

Dear Authors,

Thank you so much for making the code open source.

However, I am not able to reproduce the results for RAF to CK+, which is given in Table 1 - 72.09 % - the number corresponding to the CADA row.

This is my config file -

Log_Name='ResNet50_CropNet_withoutAFN_transferToTargetDomain_RAFtoCK+'
Resume_Model='ResNet50_CropNet_withoutAFN_trainOnSourceDomain_RAFtoCK+.pkl'
##Resume_Model=None
OutputPath='.'
GPU_ID=1
Backbone='ResNet50'
useAFN='False'
methodOfAFN='SAFN'
radius=25
deltaRadius=1
weight_L2norm=0.05
useDAN='True'
methodOfDAN='CDAN'
faceScale=112
sourceDataset='RAF'
targetDataset='CK+'
train_batch_size=32
test_batch_size=32
useMultiDatasets='False'
epochs=100
lr=0.0001
lr_ad=0.001
momentum=0.9
weight_decay=0.0001
isTest='False'
showFeature='False'
class_num=7
useIntraGCN='True'
useInterGCN='True'
useLocalFeature='True'
useRandomMatrix='False'
useAllOneMatrix='False'
useCov='False'
useCluster='False'
    
OMP_NUM_THREADS=16 MKL_NUM_THREADS=16 CUDA_VISIBLE_DEVICES=${GPU_ID} python3 TransferToTargetDomain.py \
--Log_Name ${Log_Name} \
--OutputPath ${OutputPath} \
--Backbone ${Backbone} \
--Resume_Model ${Resume_Model} \
--GPU_ID ${GPU_ID} \
--useAFN ${useAFN} \
--methodOfAFN ${methodOfAFN} \
--radius ${radius} \
--deltaRadius ${deltaRadius} \
--weight_L2norm ${weight_L2norm} \
--useDAN ${useDAN} \
--methodOfDAN ${methodOfDAN} \
--faceScale ${faceScale} \
--sourceDataset ${sourceDataset} \
--targetDataset ${targetDataset} \
--train_batch_size ${train_batch_size} \
--test_batch_size ${test_batch_size} \
--useMultiDatasets ${useMultiDatasets} \
--epochs ${epochs} \
--lr ${lr} \
--lr_ad ${lr_ad} \
--momentum ${momentum} \
--weight_decay ${weight_decay} \
--isTest ${isTest} \
--showFeature ${showFeature} \
--class_num ${class_num} \
--useIntraGCN ${useIntraGCN} \
--useInterGCN ${useInterGCN} \
--useLocalFeature ${useLocalFeature} \
--useRandomMatrix ${useRandomMatrix} \
--useAllOneMatrix ${useAllOneMatrix} \
--useCov ${useCov} \
--useCluster ${useCluster}

Kindly let me know if I am missing something.
Please help at the earliest,
Thank you again,
Megh

您好，您提供的CK+数据集中说明要对图像进行裁剪，但是我没有找到裁剪边框的数据，请问您可以提供吗

Dear authors,

ExpW does not have a validation split but it appears that you've used a validation set for evaluation. Could you please provide the val lists you used for FER2013 and ExpW?

Thanks,
Manogna

你的百度云盘预训练模型还能提供一下共享吗？

您好，请问您提供的预训练模型中GCN网络的参数是和特征提取器、分类器、域鉴别器一起训练得到的吗，如果不是的话，具体是怎样得到的呢

据我了解，ck+数据集并没有划分验证集，怎么划分呢？ 做了交叉验证？

学者您好，在您提供的CD-FER-Benchmark开源代码中，有关数据集我有一处不明白，在Utils.py里面283行有这样一个路径'/RAF/basic/Annotation/Landmarks_5/，但是在公开数据集里都没有提供Landmarks_5，请问您是怎么生成的呢？或者你怎么制作的呢？可以把您的这几个数据集的Landmarks_5发给我一份吗？或者指导一下我怎么生成这个文件，希望您有空的时候帮助解答一下我的问题，万分感谢！！！

您好，感谢您在学术上的贡献和开源代码
我想快速获得一个可以识别表情的项目，在阅读您的代码过程中发现您提供了预训练模型
所以想请问是否在完成Train on source doman后保存的模型才会拥有基本表情识别的能力？
还是加载您上传的预训练模型直接修改"TrainOnSourceDomain.py"进行test也会拥有一定的表情识别能力（只是准确率不够高）?
时间不太允许完成source doman 的训练再进行transfer,故提出这个比较蠢的疑问，十分感谢！

Source 选择AFED 数据集， Target使用 FER2013, backbone使用了ResNet18, 然而现有ICID的代码，似乎只用了Source训练集训练模型， Target 测试集准确率也只有38左右。而使用Source 和 Target 训练集一起训练模型时，Target测试集准确率达到了52.66。

您好，您提供的CK+数据集中说明要对图像进行裁剪，但是我没有找到裁剪边框的数据，请问您可以提供吗

你好，我是华南师范大学的一名学生
对于您论文的复现我想请问您几个问题：
复现时按照您的代码和数据集处理方式，从RAF-DB 到CK+ 数据集的迁移 best Accuracy 只有 78.29，无法达到85效果
您给出的docker 的image putao3/images:py3-pytorch1.3-agra 里面是python2的 ，而且没有装pytorch

Hi Authors,
Thank you so much for making the code open source. However I am facing an issue with the RAF/CK+ dataset. I downloaded the dataset from the OneDrive Link as mentioned in the repository. I am not able to find any images in the RAF dataset and CK+ dataset. Kindly tell me where can I find the images corresponding to the dataset.
I look forward to your response.
Thanks again for your help,
Megh

您好，我看了train的代码，想询问一下数据集具体是如何处理的（anno及下属文件夹等），AFE 的train_list.txt，以及各个数据集的list ，Oulu-CASIA用的是源数据集哪个文件夹的数据呢，如果可以的话，可以提供一下数据集处理相关的代码吗？

Hello, thank you very much for your work. I would like to see how the data in the Label,box and landmark files in your code are stored, ok? All I need is an example

百度网盘的链接失效，OneDrive的文件损坏

作者您好，请问可以提供已经训练好的，可以直接用于测试的AGRA模型吗，非常感谢！

你好，可以请问一下图卷积网络邻接矩阵的几个link值是根据什么确定的呢？

也就是代码中此位置的l几个link值是如何确定的呢？
非常感谢！

Good morning,

I'm trying to reproduce the results of AGRA, but I cannot match the numbers. I believe the issue is due to hyperparameter selection and data processing. Can you confirm that the hyperparameters in the TrainOnSourceDomain.sh and TransferToTargetDomain.sh files are correct?

Also, can you help me understand the data processing step you performed? I'm using the RAF-DB dataset as source and JAFFE and FER2013 as targets. From what I read in #17, I believe you split the datasets into three equal-sized parts when no train-test split is available. If you have the test split, you divide it into two equal-sized test and validation sets.

Therefore, you use 1/3 of JAFFE and 1/2 of the test set of FER2013. For the RAF-DB, on the other hand, you directly use the train split provided by the authors.

Can you confirm these processes?
Many thanks