We read every piece of feedback, and take your input very seriously.
To see all available qualifiers, see our documentation.
Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? # to your account
你好,按照DANN的理论思想,域鉴别器在训练得比较好的情况下应该是不能够很好地区分数据是来自源域还是目标域,这意味着域分类的准确率曲线应该是接近50%。但我在DANN上用自己的数据做分类实验的时候,出现了以下的问题: 1.训练过程中域分类准确率达到了100%,但在目标域上的准确率有了较大的提升; 2.源域和目标域分享同一个标签空间,即类0和类1,但即使源域在类0和类1上都达到了100%的准确率,目标域上对类0的准确率只有不到30%,对类1的准确率达到了80+%; 我想问下有没有什么方法解决这样的问题,或者说源域和目标域存在着较大的条件概率分布之类的差异,导致迁移的效果不是很好?
The text was updated successfully, but these errors were encountered:
Sorry, something went wrong.
No branches or pull requests
你好,按照DANN的理论思想,域鉴别器在训练得比较好的情况下应该是不能够很好地区分数据是来自源域还是目标域,这意味着域分类的准确率曲线应该是接近50%。但我在DANN上用自己的数据做分类实验的时候,出现了以下的问题:
1.训练过程中域分类准确率达到了100%,但在目标域上的准确率有了较大的提升;
2.源域和目标域分享同一个标签空间,即类0和类1,但即使源域在类0和类1上都达到了100%的准确率,目标域上对类0的准确率只有不到30%,对类1的准确率达到了80+%;
我想问下有没有什么方法解决这样的问题,或者说源域和目标域存在着较大的条件概率分布之类的差异,导致迁移的效果不是很好?
The text was updated successfully, but these errors were encountered: