search

yinboc / DGP

Rethinking Knowledge Graph Propagation for Zero-Shot Learning, in CVPR 2019
MIT License
320 stars 57 forks source link

Why val_loss is always 0? #5

Closed gxwangupc closed 5 years ago

gxwangupc commented 5 years ago

hi, I get val_loss=0.0000 when run "python traingcn*.py". Why is it ? Thank you.

yinboc commented 5 years ago

Thanks for your question.

Because by default the train/val split is 10/0 (see the code), that is, no validation set of nodes. This is mainly because every node is important for having a good performance when testing its neighbors.

gxwangupc commented 5 years ago

Thank you for your reply.