ProGCL
ProGCL copied to clipboard
[ICML 2022] "ProGCL: Rethinking Hard Negative Mining in Graph Contrastive Learning"
By using your parameters, it can only reach about 0.32+ acc on WikiCS. Is there anything wrong?
Thank you for your wonderful work. I encountered a problem. When running multiple experiments, do we fix the random seed and run it again, or change the random seed and...
您好,我最近想复现论文中table-5中的负采样对比方法,但是发现加上DCL和HCL结果都异常低(0.3左右)。后来使用PyGCL实现同样有个问题[https://github.com/PyGCL/PyGCL/issues/56](url),请问能否停供一下您复现DCL部分的代码,做下参考。
你好,在model.py中,更新alpha和beta参数时候,应该是p(c|s)吧,(但是fit方法里面调用responsibilities,这个r看代码计算的是p(s|(alpha,beta)),这跟论文中公式(6)不对应,谢谢
夏博士您好,我在运行您的代码的时候出现了bug,代码运行过程之中,loss = nan 经过排查,应该是between_sim = f(between_sim)写成了between_sim = f(refl_sim) 但是在修改之后,运行到batched_semi_loss_bmm函数的时候仍然出现了数据维度对不上的问题