Hello, thanks for checking this issue. I use the uploaded code and run GVB-GD on the dataset Office-31. However, I cannot reproduce the claimed results in the paper. I run " python train_image.py --gpu_id 0 --GVBG 1 --GVBD 1 --num_iterations 8004 --dset office --s_dset_path data/office/amazon_list.txt --t_dset_path data/office/dslr_list.txt --test_interval 500 --output_dir gvbgd/adn " multiple times . Then the best performace is about 89.7%. However the number in your paper is 95.0%.

Any suggestions on how to solve this ? Would you mind providing your training scripts to reproduce the results on the Office-31 ? Thank you !

Traceback (most recent call last):
File "train_image.py", line 204, in
config["out_file"] = open(osp.join(config["output_path"], "log.txt"), "w")
FileNotFoundError: [Errno 2] No such file or directory: 'office-home/san\log.txt'

Thanks for the great work.
I tried to run GVB-GD on office31 several times but cannot obtain the reported results. Can you show some training log?

Hi,could you please share the code for "Visualization of data with and without the bridge γi"?Or can you give more details about the implementation of this part?
I would really like to learn how to implement it if possible.
Thank you!

Should we make changes in the data type? Or should part of the model be written from scratch?

I am using my custom dataset with 2 classes which has 10% samples with label 1 and 90% samples with label 0 in both source and target datasets. The accuracy is quite good, and is around 98+%!. But, calling accuracy 'precision' is incorrect as precision means (no. of samples in target classified as 1)/(no. of samples in target with ground truth 1).
But, how to find precision in this case? Which part of the code must be modified to find the following:

1. sensitivity (precision) = (no. of samples in target classified as 1)/(no. of samples in target with ground truth 1).
2. specificity = (no. of samples in target classified as 0)/(no. of samples in target with ground truth 0).

We promise to give you credit in our publication. Thank you

1. 在Figure 4中，分类层中减去了γi，是怎么反向得到image的变化的，用什么工具可以做到；
2. 在公式8，D*=D1+D2，在公式10，D*=D1-D2，论文中没有对D*=D1-D2有更多的解释，代码中也是D1-D2的实现。
   我理解是D2可以表示为“domain-invariant”？