junxia97 / progcl

[ICML 2022] "ProGCL: Rethinking Hard Negative Mining in Graph Contrastive Learning"

ProGCL: Rethinking Hard Negative Mining in Graph Contrastive Learning (ICML 2022)

PyTorch implementation for ProGCL: Rethinking Hard Negative Mining in Graph Contrastive Learning accepted by ICML 2022.

Requirements

• Python 3.7.4
• PyTorch 1.7.0
• torch_geometric 1.5.0
• tqdm

Training & Evaluation

ProGCL-weight:

python train.py --device cuda:0 --dataset Amazon-Computers --param local:amazon-computers.json --mode weight

ProGCL-mix:

python train.py --device cuda:0 --dataset Amazon-Computers --param local:amazon-computers.json --mode mix

Useful resources for Pretrained Graphs Neural Networks

• The first comprehensive survey on this topic: A Survey of Pretraining on Graphs: Taxonomy, Methods, and Applications
• A curated list of must-read papers, open-source pretrained models and pretraining datasets.

Citation

@inproceedings{xia2022progcl,
  title={ProGCL: Rethinking Hard Negative Mining in Graph Contrastive Learning},
  author={Xia, Jun and Wu, Lirong and Wang, Ge and Li, Stan Z.},
  booktitle={International conference on machine learning},
  year={2022},
  organization={PMLR}
}

progcl's Issues

model.py文件之中batched_semi_loss_bmm函数似乎存在bug

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