xlearning-scu/2021-cvpr-completer

PyTorch implementation for COMPLETER: Incomplete Multi-view Clustering via Contrastive Prediction (CVPR 2021)

2021-cvpr-completer's Issues

a typo in the Autoencoder class

I guess it should be decoder_layers here ?

关于自动编码器层

你好！请问自动编码器层可以单独拿出来训练多视图的重构损失吗？

关于NoisyMNIST数据集的第一个view

查看未处理的数据集时第一个view的值全为0，请问是否是正常状态？如果不是正常状态，可否重新上传正常的该数据集？谢谢！

about baseline AE2-Nets

AE2-Nets [1] is used as a baseline in the paper. But it seems that it's not originally designed for the incomplete scenario. Why can it be used as a baseline ? How is it adapted for such scenario ?

---

1. Zhang C, Liu Y, Fu H. Ae2-nets: Autoencoder in autoencoder networks[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2019: 2577-2585.

关于CL loss与Reconstruction loss，Dual prediction loss大小差距的问题

在代码运行中，发现Reconstrucion loss，Dual prediction loss与CL loss的大小相差较大（CL loss约为其他两个loss的一千倍），请问这种设计的原因是什么？训练中是怎么解决因这种大小差距导致的优化不平衡的问题的呢？

Why is Gaussian distribution?

Dear author, I don't quite understand one point. Is this representation Z discrete or continuous? If Z is discrete, why is the variational distribution Q simulated by Gaussian distribution instead of Bernoulli distribution?

关于损失函数

您好，
我尝试用您的模型跑自己的数据集（两个视图），但是出现了对比损失为负数，且聚类指标越来越低的情况。这可能是什么原因呢？

最大化互信息和对比学习的相关实验

大佬您好，有证明说 对比学习其实也是在做最大化互信息这件事情。但是我把模型的最大化互信息替换成infoNCE后(也就是您们组里面的CC那个实例级损失函数)，发现效果没有那么惊艳，请问大佬尝试过相关实验吗？比如我在NoisyMnist数据集上尝试，发现infoNCE+重构的效果并不是很好，大概acc在76左右，而互信息+重构能够到达90+；我能想到的一些方面: 调参、归一化等问题。

about the missing rate

Hi,
I wonder, why the missing rate is halved here?

Dear author, it seems that this code can only execute two view data. Do you have a multi view version?

Dear author, it seems that this code can only execute two view data. Do you have a multi view version?

Require supplementary material

Dear author:
Hello guys, great work! Could you kindly post your supplementary material. I can't find it in Prof Peng Xi's website. I want to see the detail presentation of Eq 6.

With great thanks!