Question about the weight of DiffLoss

[Lirrr2829]

If I read the paper rightly, the DiffLoss should become larger to make the shared feature different from private feature. So maybe the beta_weight should be a negative number rather than a positive one ? Many people have noticed that the DiffLoss turns to zero quickly, and actually both the source_private_feature and the target_private_feature turn to zero, because it can make loss smaller, which is caused by the positive beta_weight. And then, the DSN model becomes the same as DANN. I tried beta_weight=-0.0001 ,the acc can sometimes reach 77.5% ,which seems not that useful......but at least the private_feature is not zero anymore.