Browse Source

circle norm bad for supervise?

new-sep-loss
Michael Pilosov, PhD 10 months ago
parent
commit
d318480b7c
  1. 4
      model.py

4
model.py

@ -58,8 +58,8 @@ class ColorTransformerModel(L.LightningModule):
alpha = self.hparams.alpha alpha = self.hparams.alpha
# N = len(outputs) # N = len(outputs)
distance = circle_norm(outputs, labels).mean() # distance = circle_norm(outputs, labels).mean()
# distance = torch.norm(outputs - labels).mean() distance = torch.norm(outputs - labels).mean()
# Backprop with this: # Backprop with this:
loss = (1 - alpha) * p_loss + alpha * distance loss = (1 - alpha) * p_loss + alpha * distance

Loading…
Cancel
Save