Open oooo111 opened 1 month ago
oooo111 您好,感谢您的提醒,我们已经注意到文中公式的错误,后续将对错误的公式进行勘误,文中公式6将修正为以下形式:
$$ l(u_i) = \log(\frac{e^{(u_i \cdot vi) / \tau}}{\sum{k \neq i} e^{(u_i \cdot vk) / \tau} + \sum{k} e^{(u_i \cdot u_k) / \tau}}) $$
再次感谢您的提醒,对于您关于公式分母和代码的问题:
cross_sim.sum(dim=1) - cross_sim.diag())
表示 cross-view similarity $\sum_{k \neq i} e^{(u_i \cdot v_k) / \tau}$ inter_sim.sum(dim=1)
表示 internal similarity $\sum_{k} e^{(u_i \cdot u_k) / \tau}$Hi oooo111,
Thank you for your reminder. We have noticed the errors in this formula mentioned in the article, and a correction will be made accordingly. Formula 6 will be revised as follows:
$$ l(u_i) = \log(\frac{e^{(u_i \cdot vi) / \tau}}{\sum{k \neq i} e^{(u_i \cdot vk) / \tau} + \sum{k} e^{(u_i \cdot u_k) / \tau}}) $$
Thank you again for your attention. Regarding your questions about the denominator in the formula and the code:
cross_sim.sum(dim=1) - cross_sim.diag())
represents the cross-view similarity $\sum_{k \neq i} e^{(u_i \cdot v_k) / \tau}$.inter_sim.sum(dim=1)
represents the internal similarity $\sum_{k} e^{(u_i \cdot u_k) / \tau}$.收到,感谢您的解答。
作者您好,请问能加您一个联系方式吗,我有点想法想请教请教你。感谢
oooo111您好,您可以通过邮箱hanshu.npu@gmail.com直接与我联系
作者你好,我注意到在论文《Graph Contrastive Learning with Adaptive Augmentation》中关于图对比学习的损失函数分母是三项相加,与您的论文stCluster公式相符,但是代码上体现的是两个损失(ui,vk)+(ui,uk),这使得我有些疑惑,请问您能给我解答下吗。感谢!
作者你好,在您的code中关于对比学习损失的复现中,其中分母是(cross_sim.sum(dim=1) + inter_sim.sum(dim=1) - cross_sim.diag())),但是在您的论文公式中分母是三项相加,请问是我的理解不对嘛。谢谢