search

Shen-Lab / GraphCL

[NeurIPS 2020] "Graph Contrastive Learning with Augmentations" by Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, Yang Shen
MIT License
548 stars 103 forks source link

Different result about Transfer Learning #52

Open StylesZhang opened 2 years ago

StylesZhang commented 2 years ago

@yyou1996 Hi! I've repeated the experiment of transfer learning using bio dataset, runing the finetune.sh by default setting. I get the result about 0.6531 and 0.6507(test easy and hard), while table 5 in your paper states that GraphCL could reach 0.6788... I'm wondering why I can't reach that score? Here is my environment: torch: 1.4.0 torch-cluster:1.5.2 torch-geometric:1.0.3 torch-scatter: 2.0.3 torch-sparse:0.5.1 torch-spline-conv: 1.2.0 I fix the function of scatter_add() of torch-geometric and delete the parameter 'fill-value', as I use the older version of torch-geometric. Otherwise it could cause Runtime Error. 'fill-value' is no longer supported for torch-scatter in version >= 2.0.0. But I guess it's not the main reason that I can't get good experiment result?

yyou1996 commented 2 years ago

Hi @HeyMercer,

Would you might take a try with lr=1e-4/1e-2 in https://github.com/Shen-Lab/GraphCL/blob/master/transferLearning_MoleculeNet_PPI/bio/finetune.sh?

StylesZhang commented 2 years ago

Hi @HeyMercer,

Would you might take a try with lr=1e-4/1e-2 in https://github.com/Shen-Lab/GraphCL/blob/master/transferLearning_MoleculeNet_PPI/bio/finetune.sh?

Thanks, I'll try!