Use euclidean distance for stop criterion

train.py

@@ -43,10 +43,6 @@ batch = 0
 # Accumulated loss for last batches
 running_loss = 0
 
-# True in the first training phase where small batches are used, and false in
-# the second phase where point positions are refined using large batches
-is_coarse_optim = True
-
 # List of training examples (not shuffled)
 classes_ordered = torch.arange(order).repeat(batch_size)
 
@@ -84,7 +80,7 @@ while total_change >= 1e-4:
     # Check for convergence
     model.eval()
     constellation = model.get_constellation()
-    total_change = (constellation - prev_constellation).abs().sum()
+    total_change = (constellation - prev_constellation).norm(dim=1).sum()
     prev_constellation = constellation
 
     # Report loss