lstm update

lstm_1b.py

@@ -28,13 +28,7 @@ class ImprovedLSTMBinaryClassifier(nn.Module):
                             dropout=dropout,
                             bidirectional=False)
         self.layer_norm = nn.LayerNorm(hidden_dim)
-        
-        # Zusätzliche Fully Connected Layers ohne ReLU
-        self.fc1 = nn.Linear(hidden_dim, 128)
-        self.fc2 = nn.Linear(128, 64)
-        self.fc3 = nn.Linear(64, 32)
-        self.fc4 = nn.Linear(32, 1)
-        
+        self.fc = nn.Linear(hidden_dim, 1)
         self.sigmoid = nn.Sigmoid()
         self.dropout = nn.Dropout(dropout)
 
@@ -43,43 +37,11 @@ class ImprovedLSTMBinaryClassifier(nn.Module):
         lstm_out = self.dropout(lstm_out)
         pooled = lstm_out[:, -1, :]  # Letztes verstecktes Zustand
         normalized = self.layer_norm(pooled)
-        
-        # Mehrere Fully Connected Schichten
-        x = self.fc1(normalized)
-        x = self.fc2(x)
-        x = self.fc3(x)
-        x = self.fc4(x)
-        
-        return self.sigmoid(x)
+        logits = self.fc(normalized)
+        return self.sigmoid(logits)
 
 # Training und Evaluation
-if __name__ == "__main__":
-    # Daten laden (Annahme: Eingebettete Daten sind bereits vorbereitet)
-    data_path = '/content/drive/MyDrive/Colab Notebooks/ANLP_WS24_CA2/data/embedded_padded'
-    train_dataset = torch.load(data_path + '/train.pt')
-    test_dataset = torch.load(data_path + '/test.pt')
-    val_dataset = torch.load(data_path + '/val.pt')
-
-    # Hyperparameter
-    input_dim = 100 
-    hidden_dim = 256
-    num_layers = 2
-    dropout = 0.3
-    batch_size = 64
-
-    # DataLoader
-    train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
-    val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False)
-    test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)
-
-    # Modell initialisieren
-    model = ImprovedLSTMBinaryClassifier(
-        input_dim=input_dim,
-        hidden_dim=hidden_dim,
-        num_layers=num_layers,
-        dropout=dropout
-    ).to(device)
-
+def train_model(model, train_loader, val_loader, test_loader, epochs=10):
     criterion = nn.BCELoss()
     optimizer = optim.Adam(model.parameters(), lr=0.001, weight_decay=1e-5)
     scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.5, patience=2, verbose=True)
@@ -88,10 +50,8 @@ if __name__ == "__main__":
     best_test_accuracy = 0
     patience = 3
     counter = 0
-    
     history = {'train_loss': [], 'val_loss': [], 'test_acc': [], 'test_f1': []}
 
-    epochs = 5
     for epoch in range(epochs):
         # Training
         model.train()
@@ -107,7 +67,7 @@ if __name__ == "__main__":
             loss = criterion(outputs, labels)
 
             loss.backward()
-            nn.utils.clip_grad_norm_(model.parameters(), 1)  
+            nn.utils.clip_grad_norm_(model.parameters(), 5)  # Gradient Clipping
             optimizer.step()
 
             total_loss += loss.item()
@@ -171,3 +131,41 @@ if __name__ == "__main__":
             if counter >= patience:
                 print("⛔ Early Stopping ausgelöst!")
                 break
+
+    return history
+
+if __name__ == "__main__":
+    # Daten laden (Annahme: Eingebettete Daten sind bereits vorbereitet)
+    data_path = '/content/drive/MyDrive/Colab Notebooks/ANLP_WS24_CA2/data/embedded_padded'
+    train_dataset = torch.load(data_path + '/train.pt')
+    test_dataset = torch.load(data_path + '/test.pt')
+    val_dataset = torch.load(data_path + '/val.pt')
+
+    # Hyperparameter
+    input_dim = 100
+    hidden_dim = 256
+    num_layers = 2
+    dropout = 0.3
+    batch_size = 64
+
+    # DataLoader
+    train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
+    val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False)
+    test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)
+
+    # Modell initialisieren
+    model = ImprovedLSTMBinaryClassifier(
+        input_dim=input_dim,
+        hidden_dim=hidden_dim,
+        num_layers=num_layers,
+        dropout=dropout
+    ).to(device)
+
+    # Training starten
+    history = train_model(
+        model,
+        train_loader,
+        val_loader,
+        test_loader,
+        epochs=5
+    )