Do you use a different loss function?

[jaelgu]

Is the loss function different from the original paper? Also the method to calculate validate score? I can't find any details about them in your thesis.

[stdio2016]

I use NT-Xent loss function, the same as the original paper. The validation score during training should be 1-second query accuracy, however it turned out to be way too high because I reused training pipeline for validation query generator. So I did not include the validation score from training process in my thesis. Instead I wrote another query generator genquery.py that more accurately simulates the query process, and only computed accuracy during testing.