Use euclidean distance for stop criterion
This commit is contained in:
		
							parent
							
								
									4f1514accf
								
							
						
					
					
						commit
						1d39184036
					
				
							
								
								
									
										6
									
								
								train.py
								
								
								
								
							
							
						
						
									
										6
									
								
								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 | ||||
|  |  | |||
		Loading…
	
		Reference in New Issue