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

[jasperhyp]

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]

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