circle norm bad for supervise?

model.py

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

newsearch.py

@@ -27,12 +27,12 @@ learning_rate_values = [1e-3]
 # learning_rate_values = [5e-4]
 
 # alpha_values = [0, .25, 0.5, 0.75, 1]  # alpha = 0 is unsupervised. alpha = 1 is supervised.
-alpha_values = [0]
+alpha_values = [1.0]
 # widths = [2**k for k in range(4, 13)]
 # depths = [1, 2, 4, 8, 16]
 widths, depths = [512], [4]
 
-batch_size_values = [64, 256, 1024]
+batch_size_values = [256]
 max_epochs_values = [100]
 seeds = list(range(21, 1992))
 optimizers = [
@@ -73,7 +73,7 @@ for idx, params in enumerate(search_params):
 python newmain.py fit \
 --seed_everything {s} \
 --data.batch_size {bs} \
---data.train_size 50000 \
+--data.train_size 0 \
 --data.val_size 10000 \
 --model.alpha {a} \
 --model.width {w} \