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
541 stars 103 forks source link

Question about the loss function in the transferLearning_MoleculeNet_PPI/chem #50

Open wwma opened 2 years ago

wwma commented 2 years ago

hello,@yyou1996, thanks for your great work! I have a question about the GraphCL in transferLearning_MoleculeNet_PPI/chem.When I run the code,the loss of each epoch is negative number. I wonder if I did something wrong，or that's what it is. thanks!

yyou1996 commented 2 years ago

Hi @wwma,

According to my understanding, loss = - torch.log(pos_sim / neg_sim), thus positive might indicate pos_sim / neg_sim > 1. It might happen when batch_size is small or augmentation is weak, or T factor is set small (e.g. <= 0.1). It looks fine to me. A simple way to turn loss positive is directly increase batch_size or lower the T factor in CL loss.