feipanir/IntraDA

目标域没有标签

Opened this issue · 7 comments

Reality-Dream7 commented

老师，如果一开始目标域有标签的话，代码是可以运行的，但是无监督域自适应不是要求目标域没有标签吗？
如果一开始目标域没有标签的话，代码就不能运行。
如果一开始目标域是有伪标签的话，那么刚开始的伪标签是怎么生成的？

feipanir commented

Hi thank you for your interests. Please prepare a validation set for your target domain. Your validation set shall contains a set of validate images and ground truth annotations.

Reality-Dream7 commented

老师，验证集val文件夹里面有标签，但是现在运行时报错，显示train文件夹里面没有标签“FileNotFoundError: [Errno 2] No such file or directory: 'E:\root\code\IntraDA\ADVENT\data\Cityscapes\gtFine\train\hamburg\hamburg_000000_087822_gtFine_labelIds.png'”

Reality-Dream7 commented

嗨，感谢您的关注。请为目标域准备验证集。您的验证集应包含一组验证图像和真实注释。

老师，验证集val文件夹里面有标签，但是现在运行时报错，显示train文件夹里面没有标签“FileNotFoundError： [Errno 2] 没有这样的文件或目录： 'E：\root\code\IntraDA\ADVENT\data\Cityscapes\gtFine\train\hamburg\hamburg_000000_087822_gtFine_labelIds.png'”

feipanir commented

I think the errors come from the loading path of your dataset. Could you double-check if your path is correctly pointed to the dataset?

Reality-Dream7 commented

我认为错误来自数据集的加载路径。您能否仔细检查您的路径是否正确指向数据集？

老师，数据集的加载路径应该是正确指向的，因为val文件夹里面有标签，而只要train文件夹里面有标签就可以执行，train文件夹里面没有标签就会报错。

feipanir commented

Thanks for the feedback. In this case, you can modify the loading function by deleting the code lines that read the labels of the train set.

bryanbocao commented

--如果一开始目标域是有伪标签的话，那么刚开始的伪标签是怎么生成的？

I have two similar questions as well.

(1) For saving the pseudo labels, which are the soft labels of pixel-wise class probabilities, should be generated and saved somewhere after the training stage from ADVENT.

(2) For loading the pseudo labels in the intra-domain training stage, they should be loaded in this line of code as labels:
https://github.com/bryanbocao/IntraDA/blob/gh-pages/intrada/train_UDA.py#L116
but what I printed seemed to be the ground truth class label from 0 to 19.

It would be appreciated if you can help with these issues @feipanir.