hlzhang109/ddg

ddg's Issues

cannot use resnet-50 model

Hi,

I encountered bugs when I try to switch from resnet-18 to resnet-50 on PACS dataset using DDG (i.e., set 'resnet18' to False in hparams_registry.py). Problem seems to appear in the AdaINGen model. Can you please upload a version that can run resnet-50 without any bugs so that we can reproduce the results in the paper? Thanks!

KeyError: 'is_ddg' && KeyError: 'is_augmix'

Sorry to bother you again and again, it's the third time I created issue :( . It seems so tough to run the code on my own device.

After preparing datasets, I try to train DDG model on my own device. When I try

python train.py —data_dir dataset —algorithm DDG —dataset PACS

it came: 'KeyError: 'is_augmix' in datasets.py, line 204, in init if hparams['is_augmix']

When I try
python train.py —data_dir dataset —algorithm DDG —dataset RotatedMNIST
it came: 'KeyError: 'is_ddg' in datasets.py, line 162, in init self.sample_pos = hparams['is_ddg']

Missing environment dependencies

Thank you for the great work!
Could you provide environment dependencies so we could run this code on our own device ? Thank you so much : )

How does this code test the trained model?

Hello, thanks for your excellent work. I find this repo only contains training commands. I can not determine how to select and test the trained model on the target domain.
I find that there is a folder called "test", do you have any reletaed commands?

Question about the gan training

Hi could you describe how you train the Gan in the training process.
I haven’t see the description in your paper's algorithm.
And your code seems too complex to read due to no comments.
It's confusing that whai is dis_update , gen_update, gan_update:)
Looking forward to your reply!

GAN training - x_ab and x_ba errors

Hi! First, congratulations on this work, it's very interesting!
I've been trying to reproduce your results (only with RotatedMNIST dataset), but then something strange happened: the pre-trained gan model that results from stage 0 generates x_ab and x_ba images wrongly, they do not present the content and style parts from the correspondent input image and end up being an image that is almost a reconstruction of the style image - a or b accordingly)... then in stage 1, these generated images continue to be retrieved with this pre-trained model as it shows in algorithms.py line 1055, so class predictions for these images will always be wrong over stage 1 training. Was this supposed to happen or did I make something wrong that you can easily identify?
Thank you very much.

No module named 'scripts.write_2images'

Thank you for the amazing work!

Could you please upload the file 'write_2images.py' in the scripts folder? When I try to train the decoder, this module is imported in algorithm.py but I couldn't find it.

Using sweep with 2 stages

Hi,
Thanks for you code!
I wanted to recreate results and run the sweep code, however you algorithm consists of 2 stage training (stage==0/1) and im not clear on how you run the sweep in this scenario. I looked in your sweep code and it is a copy of the one from DomainBed, no updates to the 2 stage training you use. Can you please share the code of the sweep?
Thanks

scripts.write_2images不存在

您好！
在scripts找不到write_2images文件，请问可以上传下嘛？

reproduce the model in rotated-mnist

Hi, could I ask for some details about reproducing the model in Rotated-MNIST?

I follow the instructions you provided, but I can't get the impressive disentangled effect just like the images in the appendix :(

why you change the backbone

I have a question, why you remove the bn layer from resnet, the removal of bn may severely influence the result, do you make fair comparison in your paper