Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

unlabeled 数据使用gt信息 #29

Closed
leonzgtee opened this issue Jun 30, 2021 · 4 comments
Closed

unlabeled 数据使用gt信息 #29

leonzgtee opened this issue Jun 30, 2021 · 4 comments
Labels
discussion Good discussion but maybe not applicable individual Individual problem and need

Comments

@leonzgtee
Copy link

你好,MI-AOD的代码中的Unlabeled set也需要从dataset到dataloader,且在epoch_based_runner.py 70-74行涉及了对gt_bboxes和gt_labels的操作,与我理解的unlabeld set应该不使用gt信息有点出入,能否帮忙指点一下。

@yuantn
Copy link
Owner

yuantn commented Jun 30, 2021

你好,这几行的操作是为了对 unlabeled set 的图像抹除定位信息,即使其所有的定位数据全部置为 -1。这样在 这里的第 479 行565 行 计算涉及到 unlabeled set 的 loss 时,就可以通过判断其定位数据是否小于 0 决定是否将该 loss 反传(即乘 0 操作)。实际上,如果 loss 没有反传,则意味着是没有使用 unlabeled set 的 GT 信息的。


Hello, these lines are to remove the localization information of the images in the unlabeled set, which is to make all of their localization information as -1. In this way, in Line 479 here and Line 565 here, when calculating the loss on the unlabeled set, we can judge whether the localization information is less than 0 to determine whether to backward propagate the loss (i.e., multiply by 0). In fact, if the loss is not back propagated, it means that the GT information of the unlabeled set is not used.

@yuantn yuantn added the discussion Good discussion but maybe not applicable label Jun 30, 2021
@leonzgtee
Copy link
Author

好的,谢谢。如果我是使用完全无标注的数据作为unlabeled set的话,需要做哪些修改呢?

@yuantn
Copy link
Owner

yuantn commented Jun 30, 2021

你可以在这些无标注数据的标注信息中添加任意一个定位框(例如,该定位框坐标可为 [0,0,0,0],类别归属为类别 1)。它们在实际训练过程中并不会被用到,因此可以任意添加。该定位框的标注格式需要和其他有标注数据的标注格式保持一致。之后,在无标注数据集的 txt 索引中加入该数据的文件名即可。


You can add any bounding box to the annotation of the unlabeled data (for example, the coordinates of the bounding box can be [0,0,0,0], and the class attribute is class 1). They will not be used in the actual training process, so they can be added arbitrarily. The annotation format of the bounding box needs to be consistent with that of other labeled data. After that, just add the file name of the unlabeled data to the txt index of the unlabeled data set.

@leonzgtee
Copy link
Author

明白了,非常感谢回复!!

@yuantn yuantn closed this as completed Jul 30, 2021
yuantn added a commit that referenced this issue Aug 16, 2021
@yuantn yuantn added the individual Individual problem and need label Sep 15, 2021
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
discussion Good discussion but maybe not applicable individual Individual problem and need
Projects
None yet
Development

No branches or pull requests

2 participants