search

mkhorr / supergluetraining

Other
13 stars 3 forks source link

Issues about the loss function #1

Closed JessieLi-RS closed 3 years ago

JessieLi-RS commented 3 years ago

Dear author,

Thank you for your contribution. However, I noticed that you have changed the original negative log-likelihood loss to another loss function. May I ask the reason??

Best, Jessie

mkhorr commented 3 years ago

Hello,

Apologies for the code’s readability, but the loss function used is still negative log likelihood. The inputs to loss_fn are the row and column indices of the P matrix (in log space) where the corresponding values are the summed terms of the loss. The terms are then weighted based on the author’s suggestion here.