In minimize training step, the training goal is to minimize the association discrepancy, the prior_loss in "loss2 = rec_loss + self.k * prior_loss" tends to 0. Therefore, the overall trend of loss2 tends to 0 too.
So why choose "score2 = -val_loss2" instead of "score2 = val_loss2" when the judgment logic is "score2 < self.best_score2 + self.delta"?
In minimize training step, the training goal is to minimize the association discrepancy, the prior_loss in "loss2 = rec_loss + self.k * prior_loss" tends to 0. Therefore, the overall trend of loss2 tends to 0 too.
So why choose "score2 = -val_loss2" instead of "score2 = val_loss2" when the judgment logic is "score2 < self.best_score2 + self.delta"?