Use best parameters as found by experimentation
This commit is contained in:
		
							parent
							
								
									5cb087d971
								
							
						
					
					
						commit
						0769a61fcf
					
				
							
								
								
									
										13
									
								
								train.py
								
								
								
								
							
							
						
						
									
										13
									
								
								train.py
								
								
								
								
							|  | @ -4,16 +4,19 @@ import torch | |||
| from matplotlib import pyplot | ||||
| from mpl_toolkits.axisartist.axislines import SubplotZero | ||||
| 
 | ||||
| torch.manual_seed(42) | ||||
| torch.manual_seed(57) | ||||
| 
 | ||||
| # Number of symbols to learn | ||||
| order = 16 | ||||
| 
 | ||||
| # Initial value for the learning rate | ||||
| initial_learning_rate = 0.1 | ||||
| 
 | ||||
| # Number of batches to skip between every loss report | ||||
| loss_report_batch_skip = 50 | ||||
| 
 | ||||
| # Size of batches | ||||
| batch_size = 32 | ||||
| batch_size = 2048 | ||||
| 
 | ||||
| # File in which the trained model is saved | ||||
| output_file = 'output/constellation-order-{}.pth'.format(order) | ||||
|  | @ -30,8 +33,8 @@ pyplot.show(block=False) | |||
| # Train the model with random data | ||||
| model = constellation.ConstellationNet( | ||||
|     order=order, | ||||
|     encoder_layers_sizes=(8, 4), | ||||
|     decoder_layers_sizes=(4, 8), | ||||
|     encoder_layers_sizes=(8, 4,), | ||||
|     decoder_layers_sizes=(4, 8,), | ||||
|     channel_model=constellation.GaussianChannel() | ||||
| ) | ||||
| 
 | ||||
|  | @ -52,7 +55,7 @@ total_change = float('inf') | |||
| 
 | ||||
| # Optimizer settings | ||||
| criterion = torch.nn.CrossEntropyLoss() | ||||
| optimizer = torch.optim.Adam(model.parameters(), lr=0.1) | ||||
| optimizer = torch.optim.Adam(model.parameters(), lr=initial_learning_rate) | ||||
| scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( | ||||
|     optimizer, verbose=True, | ||||
|     factor=0.25, | ||||
|  |  | |||
		Loading…
	
		Reference in New Issue