added grid search

BERT.py

@@ -9,6 +9,7 @@ from transformers import BertForSequenceClassification, AutoTokenizer
 import numpy as np
 from datetime import datetime
 import json
+import itertools
 
 import Datasets
 import dataset_helper
@@ -52,18 +53,26 @@ class CustomBert(nn.Module):
 if __name__ == '__main__':
     # Hyperparameter und Konfigurationen
     params = {
-        # Config
-        "max_len": 128,
         # Training
-        "epochs": 1,
+        "epochs": [1],
-        "patience": 7,
+        "patience": [7],
-        "batch_size": 32,
+        "learning_rate": [1e-5, 1e-6],
-        "learning_rate": 1e-6,
+        "weight_decay": [5e-4],
-        "weight_decay": 5e-4 ,
         # Model
-        "dropout": 0.6
+        "dropout": [0.6]
     }
 
+    # Generate permutations of hyperparameters
+    keys, values = zip(*params.items())
+    grid_params = [dict(zip(keys, v)) for v in itertools.product(*values)]
+    best_params = {}
+    best_params_rmse = -1
+    # Example usage of grid_params
+    # for param_set in grid_params:
+    #     print(param_set)
+    print('Number of grid_params:', len(grid_params))
+
+
     # Configs
     GLOVE_PATH = 'data/glove.6B.100d.txt'
     DATA_PATH = 'data/hack.csv'
@@ -72,7 +81,11 @@ if __name__ == '__main__':
     TEST_SIZE = 0.1
     VAL_SIZE = 0.1
 
-    N_MODELS = 2
+    MAX_LEN = 280
+    BATCH_SIZE = 32
+
+    N_MODELS = 1
+    USE_GIRD_SEARCH = True
 
     models = []
     timestamp = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
@@ -91,17 +104,29 @@ if __name__ == '__main__':
     print("Tokenizer Initialized")
 
     # Dataset und DataLoader
-    train_dataset = Datasets.BertDataset(tokenizer, data_split['train']['X'], data_split['train']['y'], max_len=params["max_len"])
+    train_dataset = Datasets.BertDataset(tokenizer, data_split['train']['X'], data_split['train']['y'], max_len=MAX_LEN)
-    val_dataset = Datasets.BertDataset(tokenizer, data_split['val']['X'], data_split['val']['y'], max_len=params["max_len"])
+    val_dataset = Datasets.BertDataset(tokenizer, data_split['val']['X'], data_split['val']['y'], max_len=MAX_LEN)
-    test_dataset = Datasets.BertDataset(tokenizer, data_split['test']['X'], data_split['test']['y'], max_len=params["max_len"])
+    test_dataset = Datasets.BertDataset(tokenizer, data_split['test']['X'], data_split['test']['y'], max_len=MAX_LEN)
 
-    train_loader = DataLoader(train_dataset, batch_size=params["batch_size"], shuffle=True)
+    train_loader = DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=True)
-    val_loader = DataLoader(val_dataset, batch_size=params["batch_size"], shuffle=False)
+    val_loader = DataLoader(val_dataset, batch_size=BATCH_SIZE, shuffle=False)
-    test_loader = DataLoader(test_dataset, batch_size=params["batch_size"], shuffle=False)
+    test_loader = DataLoader(test_dataset, batch_size=BATCH_SIZE, shuffle=False)
 
     subset_size = len(train_dataset) // N_MODELS
     device = ml_helper.get_device(verbose=True, include_mps=False)
 
+    # assert if N_MODLES > 1, than grid_params should be len 1
+    if N_MODELS > 1 and len(grid_params) > 1 or N_MODELS > 1 and USE_GIRD_SEARCH:
+        raise ValueError("If N_MODELS > 1, than grid_params should be len 1")
+    
+    if not USE_GIRD_SEARCH:
+        print('Using best params')
+        # load best params
+        params_name = f'models/best_params_BERT.json'
+        with open(params_name, 'r') as f:
+            best_params = json.load(f)
+        grid_params = [best_params]
+
     for i in range(N_MODELS):
         model_name = f'BERT.pt'
         hist_name = f'BERT_history'
@@ -112,7 +137,12 @@ if __name__ == '__main__':
 
             subset_indices = dataset_helper.ensemble_data_idx(train_dataset.labels, N_MODELS, i, methods='bootstrap')
             train_dataset_sub = Subset(train_dataset, subset_indices)
-            train_loader = DataLoader(train_dataset_sub, batch_size=params["batch_size"], shuffle=True)
+            train_loader = DataLoader(train_dataset_sub, batch_size=BATCH_SIZE, shuffle=True)
+
+        for para_idx, params in enumerate(grid_params):
+            if len(grid_params) > 1:
+                model_name = f'BERT_{i}_param_{para_idx}.pt'
+                hist_name = f'BERT_{i}_param_{para_idx}_history'
 
 
             model = CustomBert(dropout=params["dropout"])
@@ -153,6 +183,15 @@ if __name__ == '__main__':
             test_r2 = r2_score(test_labels, test_preds)
             print(f"Test RMSE: {test_rmse:.4f}, Test MAE: {test_mae:.4f}, Test R²: {test_r2:.4f}")
 
+            if test_rmse > best_params_rmse:
+                best_params_rmse = test_rmse
+                best_params = params
+
+        if len(grid_params) > 1:
+            best_params_name = f'models/best_params_BERT.json'
+            with open(best_params_name, 'w') as f:
+                json.dump(best_params, f)
+
 
     if N_MODELS >1:
         # Ensemble Prediction

Transformer.py

@@ -9,6 +9,7 @@ from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score
 import numpy as np
 from datetime import datetime
 import json
+import itertools
 
 import Datasets
 import dataset_helper
@@ -108,21 +109,28 @@ class TransformerBinaryClassifier(nn.Module):
 if __name__ == '__main__':
     # Hyperparameter und Konfigurationen
     params = {
-        # Config
-        "max_len": 280,
         # Training
-        "epochs": 1,
+        "epochs": [1],
-        "patience": 7,
+        "patience": [7],
-        "batch_size": 32,
+        "learning_rate": [1e-4], # 1e-4
-        "learning_rate": 1e-4, # 1e-4
+        "weight_decay": [5e-4],
-        "weight_decay": 5e-4 ,
         # Model
-        'nhead': 2, # 5
+        'nhead': [2], # 5
-        "dropout": 0.2,
+        "dropout": [0.2],
-        'hiden_dim': 2048,
+        'hiden_dim': [1024, 2048],
-        'num_layers': 6
+        'num_layers': [6]
     }
 
+    # Generate permutations of hyperparameters
+    keys, values = zip(*params.items())
+    grid_params = [dict(zip(keys, v)) for v in itertools.product(*values)]
+    best_params = {}
+    best_params_rmse = -1
+    # Example usage of grid_params
+    # for param_set in grid_params:
+    #     print(param_set)
+    print('Number of grid_params:', len(grid_params))
+
     # Configs
     GLOVE_PATH = 'data/glove.6B.100d.txt'
     DATA_PATH = 'data/hack.csv'
@@ -130,7 +138,11 @@ if __name__ == '__main__':
     TEST_SIZE = 0.1
     VAL_SIZE = 0.1
 
-    N_MODELS = 2
+    MAX_LEN = 280
+    BATCH_SIZE = 32
+
+    N_MODELS = 1
+    USE_GIRD_SEARCH = True
 
     models = []
     timestamp = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
@@ -145,17 +157,29 @@ if __name__ == '__main__':
     data_split = dataset_helper.split_data(X, y, test_size=TEST_SIZE, val_size=VAL_SIZE)
 
     # Dataset und DataLoader
-    train_dataset = Datasets.GloveDataset(data_split['train']['X'], data_split['train']['y'], word_index, max_len=params["max_len"])
+    train_dataset = Datasets.GloveDataset(data_split['train']['X'], data_split['train']['y'], word_index, max_len=MAX_LEN)
-    val_dataset = Datasets.GloveDataset(data_split['val']['X'], data_split['val']['y'], word_index, max_len=params["max_len"])
+    val_dataset = Datasets.GloveDataset(data_split['val']['X'], data_split['val']['y'], word_index, max_len=MAX_LEN)
-    test_dataset = Datasets.GloveDataset(data_split['test']['X'], data_split['test']['y'], word_index, max_len=params["max_len"])
+    test_dataset = Datasets.GloveDataset(data_split['test']['X'], data_split['test']['y'], word_index, max_len=MAX_LEN)
 
-    train_loader = DataLoader(train_dataset, batch_size=params["batch_size"], shuffle=True)
+    train_loader = DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=True)
-    val_loader = DataLoader(val_dataset, batch_size=params["batch_size"], shuffle=False)
+    val_loader = DataLoader(val_dataset, batch_size=BATCH_SIZE, shuffle=False)
-    test_loader = DataLoader(test_dataset, batch_size=params["batch_size"], shuffle=False)
+    test_loader = DataLoader(test_dataset, batch_size=BATCH_SIZE, shuffle=False)
 
     subset_size = len(train_dataset) // N_MODELS
     device = ml_helper.get_device(verbose=True, include_mps=False)
 
+    # assert if N_MODLES > 1, than grid_params should be len 1
+    if N_MODELS > 1 and len(grid_params) > 1 or N_MODELS > 1 and USE_GIRD_SEARCH:
+        raise ValueError("If N_MODELS > 1, than grid_params should be len 1")
+    
+    if not USE_GIRD_SEARCH:
+        print('Using best params')
+        # load best params
+        params_name = f'models/best_params_Transformer.json'
+        with open(params_name, 'r') as f:
+            best_params = json.load(f)
+        grid_params = [best_params]
+
     for i in range(N_MODELS):
         model_name = f'Transformer.pt'
         hist_name = f'Transformer_history'
@@ -166,7 +190,12 @@ if __name__ == '__main__':
 
             subset_indices = dataset_helper.ensemble_data_idx(train_dataset.labels, N_MODELS, i, methods='bootstrap')
             train_dataset_sub = Subset(train_dataset, subset_indices)
-            train_loader = DataLoader(train_dataset_sub, batch_size=params["batch_size"], shuffle=True)
+            train_loader = DataLoader(train_dataset_sub, batch_size=BATCH_SIZE, shuffle=True)
+
+        for para_idx, params in enumerate(grid_params):
+            if len(grid_params) > 1:
+                model_name = f'Transformer_{i}_param_{para_idx}.pt'
+                hist_name = f'Transformer_{i}_param_{para_idx}_history'
 
             # Modell initialisieren
             model = TransformerBinaryClassifier(
@@ -214,6 +243,15 @@ if __name__ == '__main__':
             test_r2 = r2_score(test_labels, test_preds)
             print(f"Test RMSE: {test_rmse:.4f}, Test MAE: {test_mae:.4f}, Test R²: {test_r2:.4f}")
 
+            if test_rmse > best_params_rmse:
+                best_params_rmse = test_rmse
+                best_params = params
+
+        if len(grid_params) > 1:
+            best_params_name = f'models/best_params_Transformer.json'
+            with open(best_params_name, 'w') as f:
+                json.dump(best_params, f)
+
     if N_MODELS >1:
         # Ensemble Prediction
         ensemble_test_preds = ml_train.ensemble_predict(models, test_loader, device)