demonph10 / mgcn Goto Github PK

作者您好，您的工作做的十分出色，我想请问一下这三个原始数据集的图片和文字应该如何获取呢？能否提供一下处理的代码呢？

Hello researchers of MGCN,
I am deeply impressed by the excellent performance of your work and am currently studying your MGCN project. I find it very insightful. However, I have a question regarding the Behavior-Aware Fuser implementation in the file mgcn.py.
In line 200, the modal embedding is calculated as follows:
side_embeds = (sep_image_embeds + sep_text_embeds + common_embeds) / 3
From my understanding, based on the original formula 15 in the paper, the modal embedding should be the sum of the shared embedding and the specific modal embeddings, without dividing by 3. Maybe like this:
side_embeds = common_embeds + (sep_image_embeds + sep_text_embeds) / 2

Could you please clarify the reasoning behind this implementation? Is this an experimental adjustment, or is there another rationale behind averaging these embeddings?
I am trying to understand the impact of this specific implementation choice on the model's performance and would greatly appreciate your insight on this matter. Thank you for your time and for sharing such an interesting project with the community. :))))))))

Best regards,
xiaodunshou

尊敬的作者，您好：

我对您的工作中，使用KDE绘制二维特征分布这部分很感兴趣，想问一下，可否寻求一份代码？
我的邮箱是: [email protected]

感谢，祝您一切顺利！