Dear author, I have a question. If i want to run train_UDA.py to train MPSCL, I need a pretrained model on source domain which you have provided, but you did not provide the code to pretrain model on source data?

I found that using the provided yaml file directly only reproduces half of the results in the paper. How can I reproduce the results reported in the paper? For example, the number of optimization iterations, batch size, etc.

Hello, Nice work. But I have a question about how to obtain the warmup model of CT2MR.

The pre-trained model could achieve ~54.1% Dice score on target test_mr. And when I loaded the warmup model and continued the warmup stage, I could obtain a ~63% Dice score on target test_mr, which was very close to your reported result of AdvEnt/AdaSeg in the MPSCL paper.

BUT, when I conducted the CT2MR warmup stage with the default config.yml from the scratch, I could only achieve ~30% Dice on target test_mr, which makes me very confused.

Could you please provide some advice? Very thanks.

Hi, thanks for the work. I have a question about the training process. When I run the training, the first 25 iterations seem to work fine, but on the 26th iteration and afterwards, I am only getting nan for all of the numerical values (for loss_seg_src_aux, loss_dice_src_aux, etc.). This seems to be caused by the model itself predicting nan.
Also, the testing process works fine, so it is only the training process that I am having the issue with.
Do you have any idea what might be causing this issue or how I can resolve it? Thank you for the help.

Hello, I'm very interested in your work. Could you explain how pre-training is conducted?

In the paper, logits are defined as the cosine similarity between the pixel feature and the anchor. However, it seems the code uses "cosine = torch.matmul(anchor_feature, class_center_feas)" directly instead of cosine similarity. Maybe I do not fully understand the paper or code. Could you explain this issue? Thanks a lot.

您好，感谢您的工作和代码，请问一下初始prototype是如何获得的，是随机的吗还是从数据集中求得的

Hi author! I have a doubt about using your code to train on another dataset. At first, I need to pre-train a segmentation model on the source domain to obtain the initial category prototypes. Is it necessary to pre-train main and aux discriminators?

Hi! I am very interested in your work, I used your code for supervised training on Cardiac CT images, my average dice result was only 86.6% at best. But the result of supervised training on Cardiac CT in your paper is 90.40%. I would like to ask if you changed the hyperparameters while doing supervised training ?

Hi, it's me again.

I want to try your code on another dataset, the abdominal dataset, which was reported in the SIFA[TMI] paper. But I cannot find the initialization code for calculating category centers. Could you plz release the corresponding code?

Thanks!

Hi, I'm pretraining the deeplabv2 to generate prototypes on another dataset. I find that the prediction is down-sampled to 33 $\times$ 33 but not up-sampled. Why is that? Did you also down-sample labels? But it seems that you did not down-sample labels before calculating BCE and dice loss in train_UDA.py.

Note that I notice that there is a label_downsample function in train_UDA.py. But it is only used in the update_class_center_iter function, and that does not affect the shape of labels used for BCE and dice loss calculation.

Thanks.

Hello, Thanks for you code, I like you work.

In domain_adaptation/eval_UDA.py line 68 :
img_mean = np.array((104.00698793, 116.66876762, 122.67891434), dtype=np.float32)
if I want to test my own dataset, I need to replace img_mean up to my dataset right?
Is the value equal to the mean value of my test dataset?

dear author, i want to train from scratch for another dataset, how can i train it, first train the warmup.yml for 4000 iterations,second using the trained parameters for training mpscl.yml?