tim-learn / atdoc Goto Github PK

Hi~ Thanks for your codes， it helps me so much to understand your paper.
Recently I use this code to run related experiments in the dataset: OFFICE-31 with the command：

python demo_uda.py --pl atdoc_na --tar_par 0.2 --dset office --max_epoch 100 --s 0 --t 1 --gpu_id 0 --method DDC --output logs/uda/run1/

However, the experiment result of office-31 in A to D cannot achieve 96% in your paper but got 89%， which made me so confused.
I have tried to revise the batch size to 16 or 28，but does not work.
(Cause of my GPU memory limitation, I cannot set the batch size to 32 or 36 )
Can you share more detail of the hyper-parameter for this dataset training, I will be very grateful.

I simply use the demo of 【'ATDOC-NA' combined with 'CDAN+E'】,change the dataloader with mine and set epochs=10,

python demo_uda.py --pl atdoc_na --tar_par 0.2 --dset VISDA-C --max_epoch 10 --s 0 --t 1 --gpu_id 1 --method CDANE --output logs/uda/run1/

but result only achieve 69.18%.
Can you relese hyper-parameters of training on VISDA-C dataset?
Thanks!

Hi,thanks for sharing your code! I have a question about calculating the cosine distance in demo_uda , line 267. dis = -torch.mm(features_target.detach(), mem_fea.t())I think before calculating the cosine distance,the features should be normalized.May be I'm missing something?

Is there a big difference in performance between the fully-connected layer often used in image classification tasks compared to the Nearest centroid Classifier used here?

Hi,thanks for sharing your code! I have a question in demo_uda , line 169. dis[di, idx[di]] = torch.max(dis) What is the form of idx, I can't get it through iter.next() in the dataset I built myself (not the image public dataset)，thank you

Thanks for sharing the code~ After reading the codebase, I find the provided demo is aimed at unsupervised DA, while your paper also provide some promising results on SSDA setting. Therefore I wonder if there are any difference between hyperparameter settings under DA and SSDA scenario.

For example

• the training batch distribution for label-source, label-target and unlabel-target
• whether the labeled target feature & score are used to update memory
• the scheme to adjust weight $lambda$ of self-supervised loss term
• whehter the hyperparameters are different for other datasets like DomainNet

BTW, I find the following code for NA classifier (Line 305 from demo_uda.py) seems not consistent with the Eq (6) in paper

outputs_target = softmax_out**2 / ((softmax_out**2).sum(dim=0))

From Eq (6) in the paper, the normalization is over prediction on all classes to make the distribution sharper, while in the code the parameter is dim=0, resulting normalization over all unlabeled data samples, will the normalization direction make big difference in final results ?

Appreciation if you can provide more detail about it~