Fixed bug

Author: david-baek

notebooks/case_study_circle.ipynb

@@ -40,7 +40,7 @@
 lr = 0.002
 weight_decay = 0.1
 
-n_exp=1
+n_exp=4
 
 param_dict = {
     'seed': seed,

src/run_exp.py

@@ -120,7 +120,7 @@ def train_single_model(param_dict: dict):
     test_dataloader = torch.utils.data.DataLoader(test_dataset, batch_size=batch_size, shuffle=False)
 
     ret_dic = {}
-    ret_dic["results"] = model.train(param_dict={'num_epochs': num_epochs, 'learning_rate': lr, 'weight_decay':weight_decay, 'train_dataloader': train_dataloader, 'test_dataloader': test_dataloader, 'device': device, 'video': video, 'verbose': verbose, 'lambda': lamb_reg})
+    ret_dic["results"] = model.train(param_dict={'model_id':model_id,'num_epochs': num_epochs, 'learning_rate': lr, 'weight_decay':weight_decay, 'train_dataloader': train_dataloader, 'test_dataloader': test_dataloader, 'device': device, 'video': video, 'verbose': verbose, 'lambda': lamb_reg})
     ret_dic["model"] = model
     ret_dic["dataset"] = dataset
 

src/utils/driver.py

@@ -25,6 +25,7 @@ def train(self, param_dict: dict):
         test_dataloader = param_dict['test_dataloader']
         device = param_dict['device']
         weight_decay = param_dict['weight_decay']
+        model_id = param_dict['model_id']
         video = False if 'video' not in param_dict else param_dict['video']
 
         verbose = True
@@ -61,14 +62,16 @@ def train(self, param_dict: dict):
                 batch_inputs = batch_inputs.to(device)
                 batch_targets = batch_targets.type(torch.LongTensor).to(device)
                 optimizer.zero_grad()
-                logits = self.forward(batch_inputs)
+                outputs = self.forward(batch_inputs)
 
 #               class_counts = torch.bincount(batch_targets.squeeze(), minlength=self.vocab_size).double() + 1e-8
 #               class_weights = 1 / class_counts.cuda()
 
                 criterion = nn.CrossEntropyLoss()#weight=class_weights)
-                
-                loss = criterion(logits, batch_targets.squeeze())
+                if 'H_' in model_id: # Harmonic Model
+                    loss = (-1)*(outputs[torch.arange(outputs.size(0)), batch_targets.squeeze()].mean())
+                else:
+                    loss = criterion(outputs, batch_targets.squeeze())
 
                 if hasattr(self.embedding, 'weight'):
                     total_loss = loss + lamb_reg * torch.mean(torch.sqrt(torch.mean(self.embedding.weight**2, dim=0)))
@@ -80,7 +83,7 @@ def train(self, param_dict: dict):
                 train_loss += loss.item()
 
                 # Compute training accuracy
-                _, predicted = torch.max(logits, 1)
+                _, predicted = torch.max(outputs, 1)
                 train_correct += (predicted == batch_targets).sum().item()
                 train_total += batch_targets.size(0)