Hello,

1. When I run the code of PDA on my own dataset, it appears that the loss of dcc is negative. I've read the appendix A, which describes that the loss of dcc represents the inter-class domain discrepancy and intra-class domain discrepancy. If it means that the inter-class domain discrepancy is large, leading to the loss of dcc is negative. How can I fix it? Is it possible to set hyper parameters to fix it?
2. In the answer of #6 issuse, you replied that num_pclass (A) * num_samples (B) are hypers for class-wise sampling in CDD, it enables each batch composes of num_class* num_samples samples, i.e., contain samples from B classes, and each class contains A samples. So I'm wondering that the num_class and num_samples is set according to the source data or target data?

Thanks for your reading. I'm looking forward to your reply.

Hello! I wanted to run your method but I have this error:

$ CUDA_VISIBLE_DEVICES=0 python office_run.py note=replicating_r0 setting=osda source=amazon target=dslr
Traceback (most recent call last):
  File "office_run.py", line 8, in <module>
    from init_config import *
  File "/home/francesco/projects/Domain-Consensus-Clustering/init_config.py", line 4, in <module>
    from utils.flatwhite import *
  File "/home/francesco/projects/Domain-Consensus-Clustering/utils/__init__.py", line 1, in <module>
    from .cdd import *
  File "/home/francesco/projects/Domain-Consensus-Clustering/utils/cdd.py", line 2, in <module>
    from .dist  import to_cuda
ModuleNotFoundError: No module named 'utils.dist'

Indeed there is not utils/dist.py file in the repo, could you upload it?

Hi @Solacex I found a calculation mistake in Table 5 Row4. The average for Office-Home is coming out as 63.01 whereas it's written as 64.2 in the paper. Kindly rectify this.

I wanted to ask if we should use the correct average when we cite your paper in our new work?

Thanks!

According to the paper, Office-31 OSDA class slipt setting is 10/0/11, and the classes are separated according to their alphabetical order. But it seems like there is no released config file for this part, so I am quite confused whether target private 11 refers to class20-30 or class10-20? Many thanks.

Why do you use the labels of target domain (i.e. t_label) to calculate loss during training models?I think the use of 't_label' is forbidden for unsupervised learning tasks. PS: .\trainer\dcc1_trainer.py: class Trainer(BaseTrainer): def iter(self, i_iter): ...... en_loss = self.memory.forward(t_feat, t_label, self.config.t, False)

Hi, i read your code and paper. In the paper, you show the results HM (%) on Office under the OSDA scenario (ResNet-50). However, in the code, the class setting of Office is inconsistent with the previous method such as STA and ROS. In the previous method, the setting of office-31 is 10 known classes and 11 unknown classes. However, in the code, the unknown classes is the left 21 classes. Could u please explain it? Thanks!

Hi @Solacex

Can you please tell me what "Acc." refers to in table 1. It would be great if you could tell how to calculate that exactly. Is it OS or OS* as defined in prior works or is it something else?

Thanks!

Sorry to bother you. I have the same question as the previous issue. Which result you report in your paper, the result of the last step or the best step？ Thank you for your reply.
Best wishes.

Thanks a lot for open-sourcing such a wondeful work!

I have a question about the final results. Which result you report in your paper, the result of the last step or the best step.

Look forward to your reply. Thanks in advance.

Thanks for sharing your code! In Prerequisites section, you suggest input this command :pip install -r requirements.txt to install any toolkits needed. However, I could not find requirements.txt this file in any folder. Where could I find this txt file?

Many Thanks

Where does self.s_loader (line 19), self.loader (line 37), self.t_loader (line 58) in Trainer (dcc1_trainer.py) come from ? In BaseTrainer, you set self.test_loader, self.src_loader, self.tgt_loader. Does these loaders correspond to the loaders in Trainer?

Hi Team,

Thanks for sharing the great work!

I've noticed in offcehome_run.py, CSDA is not provided.

https://github.com/Solacex/Domain-Consensus-Clustering/blob/5a626e1a7b294ed8fa71205bbc549a93b566e864/offcehome_run.py#L28 After I added: elif config.setting=='cda': config.cls_share = 65 config.cls_src = 0 config.cls_total = 65

I noticed there is a huge gap between the result and paper you reported in support material Table C. here is the result I've got: Ar->Pr[best]: 0.6168955022899004 0.6168955022899004 [H-Score]: 0.0 0.0 0.0 0.0 0.0 Ar->Pr[last]: 0.3347919933497906 0.3347919933497906 [H-Score]: 0.0 0.0 0.0 Ar->Cl[best]: 0.4183573762957866 0.4183573762957866 [H-Score]: 0.0 0.0 0.0 0.0 0.0 Ar->Cl[last]: 0.3801841050529709 0.3801841050529709 [H-Score]: 0.0 0.0 0.0 Ar->Re[best]: 0.6801177620887756 0.6801177620887756 [H-Score]: 0.0 0.0 0.0 0.0 0.0 Ar->Re[last]: 0.6801177620887756 0.6801177620887756 [H-Score]: 0.0 0.0 0.0 Pr->Ar[best]: 0.4206200426587692 0.4206200426587692 [H-Score]: 0.0 0.0 0.0 0.0 0.0 Pr->Ar[last]: 0.3828089099090833 0.3828089099090833 [H-Score]: 0.0 0.0 0.0 Pr->Cl[best]: 0.3435283417598559 0.3435283417598559 [H-Score]: 0.0 0.0 0.0 0.0 0.0 Pr->Cl[last]: 0.29526575485674234 0.29526575485674234 [H-Score]: 0.0 0.0 0.0 Pr->Re[best]: 0.6622800114636238 0.6622800114636238 [H-Score]: 0.0 0.0 0.0 0.0 0.0 Pr->Re[last]: 0.6590736272243353 0.6590736272243353 [H-Score]: 0.0 0.0 0.0 Cl->Ar[best]: 0.44268233615618485 0.44268233615618485 [H-Score]: 0.0 0.0 0.0 0.0 0.0 Cl->Ar[last]: 0.32432382525159764 0.32432382525159764 [H-Score]: 0.0 0.0 0.0 Cl->Pr[best]: 0.5993400445924355 0.5993400445924355 [H-Score]: 0.0 0.0 0.0 0.0 0.0 Cl->Pr[last]: 0.5993400445924355 0.5993400445924355 [H-Score]: 0.0 0.0 0.0 Cl->Re[best]: 0.5868140732009823 0.5868140732009823 [H-Score]: 0.0 0.0 0.0 0.0 0.0 Cl->Re[last]: 0.5521672795311763 0.5521672795311763 [H-Score]: 0.0 0.0 0.0 Re->Ar[best]: 0.5438899568067147 0.5438899568067147 [H-Score]: 0.0 0.0 0.0 0.0 0.0 Re->Ar[last]: 0.48911217634494486 0.48911217634494486 [H-Score]: 0.0 0.0 0.0 Re->Pr[best]: 0.7569770986070999 0.7569770986070999 [H-Score]: 0.0 0.0 0.0 0.0 0.0 Re->Pr[last]: 0.7569770986070999 0.7569770986070999 [H-Score]: 0.0 0.0 0.0 Re->Cl[best]: 0.44347345402034427 0.44347345402034427 [H-Score]: 0.0 0.0 0.0 0.0 0.0 Re->Cl[last]: 0.44347345402034427 0.44347345402034427 [H-Score]: 0.0 0.0 0.0

And here is result in the SM you've reported:

Except for this CSDA experiments, the other experiments go well, though a little unstable.

Most of my packages are newer than you.Maybe there won't be a big problem,My video memory is only 8GB,I hope I can finish this project，I hope you can help me then，thanks!

By looking at the config files for Office and OfficeHome exps I noticed that one of the difference is in the values of num_pclass and num_samples parameters for data loading. How are these params used? How should I choose their values if testing the code on a different dataset?

On the one hand, I wonder if you could offer the training code on the VisDA dataset. On the other hand, there are still some mistakes in this project, as the lack of ./utils/dist.py. Looking forward to your reply~