search

illidanlab / MoCL-DK

Implementation for the paper MoCL: Contrastive Learning on Molecular Graph with multi-level Domain Knowledge
GNU General Public License v3.0
40 stars 9 forks source link

Why not perform the transfer learning setting? #5

Open ha-lins opened 3 years ago

ha-lins commented 3 years ago

Hi @mengyings,

Thanks for the great work! I have a minor question as the title shows. You say: We do not evaluate transfer learning setting in this paper where a pretrained model is applied to another dataset. The transfer learning setting could be more challenging and practical. Following Hu et. al, you can pretrain on the unlabeled ZINC15 database, and then evaluated on these downstream tasks. Could it be possible to perform this setting for MoCL-DK? Thanks!

Reference: [1] Strategies to pre-train graph neural networks. Hu et al. 2020.

mengyings commented 3 years ago

Yes, we are currently pretraining on the ZINC dataset and will release the pre-trained model once done.