您好，我想用您的训练后的权重复现一下您的实验结果，您公布的是百分之2和百分之5训练数据的权重，按道理说应该是M&MS数据集的啊，但为什么最后经过证明我认为是SCGM数据集的权重。SCGM是用的百分之20的数据，所以这个权重到底是哪个数据集的呢？可以解答一下吗？

您好，我看了您的代码，在训练和测试的代码中您都设置了种子数，但是对于训练代码，用同一个种子数的情况下重复多次实验得到的不同权重进行测试，最终得到的测试结果是不同的，我想问一下，为什么设置了种子数但是多次训练得到的结果不同呢？我猜想是不是这个原因：upsample_bilinear2d_backward_out_cuda does not have a deterministic implementation，另外，您在论文中得到的实验结果是多次实验得到的平均值吗？

Hello, I would like to ask how the M&M dataset is downloaded? I found the designated website, but I ran into some difficulties. Is there any other way to download?

Hello, I see that your code comparison articles are all based on unet as the backbone, but your article is based on resnet50 as the backbone. Is it unfair to compare with different backbones?

您好，我想问一下我下载完代码后，用SCGM数据集训练时，出现loss都为Nan的情况，是什么原因造成的呢？

Hello, my current focus is also on semi supervised domain generalization, and I am very interested in your paper. Note that your paper refers to the Meta article when processing data, but in the publicly available code for processing the MMs dataset, the length of labeled data in each training domain is aligned, while in your paper, the length of labeled and unlabeled data is aligned, which increases the amount of data to some extent. Is this comparison fair and reasonable?

Hello, Dr. Yao. I would like to ask in your EPL_SemiDG code, when saving the best checkpoint, your code is to save it if val_dice > best_dice. I would like to ask why you don't choose if test_dice > best_dice to save it? If you have time can you reply? Thank you so much!

你好，实验中需要在哪里修改标记数据比例呢？我之前做的半监督医学图像分割，是将batch_size前一半数据从标记数据选取，后一半从无标记数据选取 ，但是半监督域泛化好像并不是这样处理的。

您好，我想问下，只用有标签数据和有标签数据无标签数据都用这两者在您的实验中效果差别大吗？我跑了您给的代码，设定了种子数，我对上面这两种情况进行了实验，发现在SCGM数据集的D域中两者测试准确率只相差0.04，在A域的实验中有标签数据无标签数据都用比只用有标签数据的效果还差。我想问下这是什么原因呢？

您好，图表中的标准偏差是怎么计算的啊？

This code is a great implementation of the research article that you published. I read through the code, and it is easy to read/understand, but I have three inquiries.

1. Are the unique images in the training set the same as in the validation set with no split?
   because I can see "val_dataset." is equal to the "label_dataset" before you duplicate the "label_dataset" multiple times to "new_label_dataset."

2. if the first question is yes, then are you saving the model every time it surpasses itself on the "val_dataset." which the model already sees during training?

3. The results published in the article are that after early stopping of the model and blindly evaluating it on the test set directly or is it the best dice score that the model output throughout the training process?

Thank you so much,