hzhao98/gdcl

Graph Debiased Contrastive Learning with Joint Representation Clustering

Graph Debiased Contrastive Learning with Joint Representation Clustering

This is the pytorch implementation of Graph Debiased Contrastive Learning (GDCL) model that is published in IJCAI 2021. (https://www.ijcai.org/proceedings/2021/0473.pdf).

Requirements

Python 3.7
Pytorch 1.2.0

Reference

@inproceedings{zhao2021graph,
  title={Graph debiased contrastive learning with joint representation clustering},
  author={Zhao, Han and Yang, Xu and Wang, Zhenru and Yang, Erkun and Deng, Cheng},
  booktitle={Proc. IJCAI},
  pages={3434--3440},
  year={2021}
}

gdcl's Issues

作者你好，请问可以上传详细的超参数列表吗？我无法复现出论文的效果

Hello, may I know the version of dgl？

Reproduction problem

Hi @hzhao98,

Thanks for your great work.

I have difficulty reproducing the code by getting the following error.

File "train.py", line 384, in train_ucgl
  pos_sam_id = random.sample(range(0, class_pos.shape[0]), int(pos_size))
File "/home/.conda/envs/torch18/lib/python3.8/random.py", line 363, in sample
  raise ValueError("Sample larger than population or is negative")
ValueError: Sample larger than population or is negative

I have already adopted the parameters in train.py and used the pre-trained models in this repo. Please let me know how to reproduce the results in the paper, thanks a lot.

模型架构问题

您好！我拜读了您的论文与代码，我发现论文中的模型架构与代码中的好像不太一致。
1、gcn1的权重好像是随机初始化的并且没有更新，也不共享gcn2的参数。
2、正样本是否只用的数据增强后的同一簇中的实例（样本）？
3、请问您是如何做的预训练？
非常抱歉打扰您，期待您的回复。

求助帖

作者大大好，
请问能提供一下你预训练好的文件吗？
我还想请教一下用MVGRL预训练
谢谢大大Orz