Changing supervised regularization term (alpha) from scalar to balanced weight
Opened this issue · 0 comments
stompsjo commented
It may be advantageous to change the current implementation of alpha from a scalar directly applied to the labeled contrastive loss term to a balanced weight that works on both versions. i.e. changing this line:
RadClass/models/PyTorch/lightModel.py
Line 242 in 093040a
To this:
curr_loss['loss']['losses'][indices] *= self.alpha
curr_loss['loss']['losses'][~indices] *= (1-self.alpha)