Formatting

src/hyper_search.py

@@ -71,7 +71,7 @@ def parse_args():
     parser.add_argument('--ema_decay', default=0.999, type=float,
                         help='EMA decay rate')
     parser.add_argument('--total_hour', default=50, type=int, help='total hours to run')
-    
+
     parser.add_argument('--train_dir', help='directory to output the result')
 
     args = parser.parse_args()
@@ -126,18 +126,19 @@ def train(args):
     precalculated_class_weights = config[args.dataset_name]['class_weights']
     weights = torch.Tensor(precalculated_class_weights)
     weights = weights.to(args.device)
-    
+
     model = get_model(args)
     model = model.to(args.device)
-    
+
     optimizer = get_optimizer(args)
     train_loader, unlabel_loader, val_loader, test_loader = get_dataloaders(args)
-      
+
     writer = SummaryWriter(args.train_dir)
-   
+
     # Initialize scheduler based on args
     if args.lr_schedule_type == 'CosineLR':
-        scheduler = get_cosine_schedule_with_warmup(optimizer, args.lr_warmup_epochs, args.lr_cycle_epochs)
+        scheduler = get_cosine_schedule_with_warmup(
+            optimizer, args.lr_warmup_epochs, args.lr_cycle_epochs)
     elif args.lr_schedule_type == 'FixedLR':
         scheduler = get_fixed_lr(optimizer, args.lr_warmup_epochs, args.lr_cycle_epochs)
     else:
@@ -149,33 +150,34 @@ def train(args):
     args.start_epoch = 0
     current_count = 0
     total_time = 0
-    
+
     # Early stopping
-    early_stopping = EarlyStopping(patience=args.patience, initial_count=current_count)       
-    
+    early_stopping = EarlyStopping(patience=args.patience, initial_count=current_count)
+
     start_time = time.time()
-    
+
     for epoch in range(args.start_epoch, args.train_epoch):
         # Train
-        train_losses = train_one_epoch(args, weights, train_loader, model, optimizer, scheduler, epoch)
-
+        train_losses = train_one_epoch(args, weights, train_loader,
+                                       model, optimizer, scheduler, epoch)
+
         # Evaluate
         val_loss, val_acc, val_labels, val_preds = eval_model(args, val_loader, model, epoch)
         test_loss, test_acc, test_labels, test_preds = eval_model(args, test_loader, model, epoch)
-        
+
         # Update best scores
         is_best = val_acc > best_val_acc
         if is_best:
             best_val_acc = val_acc
             best_test_acc = test_acc
-            
+
         # Log metrics
         writer.add_scalar('train/loss', np.mean(train_losses), epoch)
         writer.add_scalar('val/accuracy', val_acc, epoch)
         writer.add_scalar('val/loss', val_loss, epoch)
         writer.add_scalar('test/accuracy', test_acc, epoch)
         writer.add_scalar('test/loss', test_loss, epoch)
-        
+
         # Save checkpoint
         save_checkpoint({
             'epoch': epoch + 1,
@@ -185,31 +187,29 @@ def train(args):
             'optimizer': optimizer.state_dict(),
             'scheduler': scheduler.state_dict(),
         }, is_best, args.train_dir)
-        
+
         # Early stopping check
         if early_stopping(val_acc):
             print(f'Early stopping triggered after epoch {epoch}')
             break
-
-
+
         # Time tracking
         epoch_time = time.time() - start_time
         total_time += epoch_time
         start_time = time.time()
 
-
     # Save final summary
     summary = {
         'best_val_accuracy': best_val_acc,
         'best_test_accuracy': best_test_acc,
         'total_epochs': epoch + 1,
         'total_time': total_time
     }
-    
+
     with open(os.path.join(args.train_dir, 'training_summary.json'), 'w') as f:
         json.dump(summary, f)
-        
-    writer.close()   
+
+    writer.close()
     return best_val_acc, best_test_acc
 
 

src/utils.py

@@ -12,8 +12,8 @@ def sample_loguniform(low=0, high=1, size=1, coefficient=1, base=10):
     return coefficient*np.power(base, power_value)
 
 
-def sample_uniform(low=0.0, high=1.0, size=1, decimal=1):  
-    return round(np.random.uniform(low=low, high=high, size=size)[0],1)
+def sample_uniform(low=0.0, high=1.0, size=1, decimal=1):
+    return round(np.random.uniform(low=low, high=high, size=size)[0], 1)
 
 
 def str2bool(s):
@@ -23,25 +23,25 @@ def str2bool(s):
         return False
     else:
         raise NameError('Bad string')
-    
+
 
 def save_checkpoint(state, is_best, checkpoint, filename='last_checkpoint.pth.tar'):
     filepath = os.path.join(checkpoint, filename)
     torch.save(state, filepath)
     if is_best:
         shutil.copyfile(filepath, os.path.join(checkpoint,
                                                'model_best.pth.tar'))
-        
+
 
 def set_seed(seed):
     random.seed(seed)
     np.random.seed(seed)
     torch.manual_seed(seed)
-    
-      
+
+
 def get_cosine_schedule_with_warmup(optimizer,
                                     lr_warmup_epochs,
-                                    lr_cycle_epochs, #total train epochs
+                                    lr_cycle_epochs,  # total train epochs
                                     num_cycles=7./16.,
                                     last_epoch=-1):
     def _lr_lambda(current_epoch):
@@ -51,10 +51,10 @@ def _lr_lambda(current_epoch):
             float(max(1, float(lr_cycle_epochs) - lr_warmup_epochs))
         return max(0., math.cos(math.pi * num_cycles * no_progress))
 
-    return LambdaLR(optimizer, _lr_lambda, last_epoch)    
+    return LambdaLR(optimizer, _lr_lambda, last_epoch)
 
 
 def get_fixed_lr(optimizer, lr_warmup_epochs, lr_cycle_epochs, num_cycles=7./16., last_epoch=-1):
     def _lr_lambda(current_epoch):
         return 1.0
-    return LambdaLR(optimizer, _lr_lambda, last_epoch)    
+    return LambdaLR(optimizer, _lr_lambda, last_epoch)