search

zjunlp / OntoProtein

[ICLR 2022] OntoProtein: Protein Pretraining With Gene Ontology Embedding
MIT License
141 stars 22 forks source link

Exploding KE loss when training a new `protein-go` set #26

Closed jasperhyp closed 1 year ago

jasperhyp commented 1 year ago

Hi! I was trying to re-train your model with a new set of protein-GO relations, and it looks like the KE losses for positive and negative samples are exploding unexpectedly. After a few steps, the positive one increases to several tens of thousands and the negative one decreases to minus several tens of thousands. It seems they would continue to explode. Have you encountered this issue in your experiments? I'm thinking of confining the score to [0, 1] before inputting into the KE loss. Thanks in advance！

Alexzhuan commented 1 year ago

Sorry, we have not encountered the above issue in our experiments.