train: augmented model

results/config_snapshot_20250718_001955.py

@@ -0,0 +1,58 @@
+class Config:
+    # Models to train
+    MODELS_TO_TRAIN = ["MobileNetV2"]
+
+    # Data parameters
+    IMG_SIZE = (224, 224)
+    BATCH_SIZE = 32
+    TRAIN_SPLIT = 0.80
+    VALIDATION_SPLIT = 0.25
+
+    # Training parameters
+    INITIAL_EPOCHS = 20
+    FINE_TUNE_EPOCHS = 10
+    INITIAL_LEARNING_RATE = 0.0005
+    FINE_TUNE_LEARNING_RATE = 0.0001
+
+    # Model parameters
+    NUM_CLASSES = 23
+    DROPOUT_RATE = 0.5
+    NUM_FROZEN_LAYERS = 2  # Number of layers to freeze in feature extraction
+
+    # Enhanced Architecture Features (set to True to enable)
+    USE_ENHANCED_ARCHITECTURE = False  # Multi-branch architecture with attention
+    USE_PROGRESSIVE_UNFREEZING = False  # Progressive unfreezing strategy
+    USE_ADVANCED_AUGMENTATION = False  # Advanced data augmentation
+    USE_ENSEMBLE_LEARNING = False  # Ensemble learning
+    USE_TEST_TIME_AUGMENTATION = False  # Test-time augmentation
+
+    # Progressive unfreezing parameters
+    PROGRESSIVE_PHASES = 3  # Number of progressive unfreezing phases
+
+    # Ensemble learning parameters
+    ENSEMBLE_STRATEGY = 'soft_voting'  # 'soft_voting', 'hard_voting', 'weighted_voting', 'stacking'
+    ENSEMBLE_MODELS = ["MobileNetV2", "MobileNetV3Large", "MobileNetV3Small"]
+    META_LEARNER_EPOCHS = 50
+
+    # Test-time augmentation parameters
+    TTA_AUGMENTATIONS = 5  # Number of augmentations per test image
+
+    # Advanced training parameters
+    USE_MIXUP = False  # Mixup data augmentation
+    USE_CUTMIX = False  # CutMix data augmentation
+    USE_LABEL_SMOOTHING = False  # Label smoothing
+    LABEL_SMOOTHING_FACTOR = 0.1
+
+    # Learning rate scheduling
+    USE_COSINE_ANNEALING = False  # Cosine annealing scheduler
+    USE_WARM_RESTARTS = False  # Warm restarts
+
+    # Regularization
+    USE_DROPOUT_SCHEDULING = False  # Adaptive dropout scheduling
+    USE_WEIGHT_DECAY = False  # L2 weight decay
+    WEIGHT_DECAY_FACTOR = 1e-4
+
+    # Paths
+    DATA_DIR = "data"
+    MODEL_SAVE_DIR = "saved_models"
+    RESULTS_DIR = "results"

results/model_comparison_20250718_001955.csv

@@ -0,0 +1,2 @@
+,Accuracy (%),Loss,Size (MB),Parameters
+MobileNetV2,49.59,1.6897,9.3222,2287447

results/training_history_20250718_001955.csv

@@ -0,0 +1,31 @@
+model,epoch,train_accuracy,val_accuracy,train_loss,val_loss
+MobileNetV2_feature_extraction,1,0.10285714268684387,0.1804979294538498,3.2679028511047363,2.7532925605773926
+MobileNetV2_feature_extraction,2,0.22207792103290558,0.28215768933296204,2.6492979526519775,2.5107262134552
+MobileNetV2_feature_extraction,3,0.32181817293167114,0.3381742835044861,2.3156282901763916,2.323220729827881
+MobileNetV2_feature_extraction,4,0.37584415078163147,0.3838174343109131,2.1061758995056152,2.228482484817505
+MobileNetV2_feature_extraction,5,0.4171428680419922,0.3921161890029907,1.9495772123336792,2.1282544136047363
+MobileNetV2_feature_extraction,6,0.4527272582054138,0.4066390097141266,1.840901255607605,2.0412991046905518
+MobileNetV2_feature_extraction,7,0.47220778465270996,0.4273858964443207,1.7568695545196533,1.970870852470398
+MobileNetV2_feature_extraction,8,0.5070129632949829,0.4585062265396118,1.6532152891159058,1.9253556728363037
+MobileNetV2_feature_extraction,9,0.5210389494895935,0.46265560388565063,1.5946297645568848,1.8834987878799438
+MobileNetV2_feature_extraction,10,0.5425974130630493,0.47095435857772827,1.5383957624435425,1.8493201732635498
+MobileNetV2_feature_extraction,11,0.5488311648368835,0.46265560388565063,1.49072265625,1.7981990575790405
+MobileNetV2_feature_extraction,12,0.5703896284103394,0.4605809152126312,1.4195979833602905,1.7736471891403198
+MobileNetV2_feature_extraction,13,0.5807791948318481,0.4564315378665924,1.4124072790145874,1.7660542726516724
+MobileNetV2_feature_extraction,14,0.5774025917053223,0.46887966990470886,1.3677922487258911,1.7547590732574463
+MobileNetV2_feature_extraction,15,0.6015584468841553,0.46680498123168945,1.3222650289535522,1.734666347503662
+MobileNetV2_feature_extraction,16,0.6051948070526123,0.46473029255867004,1.3051408529281616,1.7301479578018188
+MobileNetV2_feature_extraction,17,0.6038960814476013,0.4730290472507477,1.2859197854995728,1.7186366319656372
+MobileNetV2_feature_extraction,18,0.6283116936683655,0.4792531132698059,1.2456966638565063,1.712315559387207
+MobileNetV2_feature_extraction,19,0.6168830990791321,0.4792531132698059,1.2730422019958496,1.7302953004837036
+MobileNetV2_feature_extraction,20,0.621558427810669,0.4792531132698059,1.2226208448410034,1.705318808555603
+MobileNetV2_fine_tuned,1,0.635064959526062,0.48547717928886414,1.1813465356826782,1.7019144296646118
+MobileNetV2_fine_tuned,2,0.6475324630737305,0.48755186796188354,1.172446846961975,1.6987230777740479
+MobileNetV2_fine_tuned,3,0.6496104001998901,0.4834024906158447,1.1667516231536865,1.7070080041885376
+MobileNetV2_fine_tuned,4,0.6407791972160339,0.4958506226539612,1.1737589836120605,1.6940451860427856
+MobileNetV2_fine_tuned,5,0.6376623511314392,0.48962655663490295,1.1516849994659424,1.7008932828903198
+MobileNetV2_fine_tuned,6,0.6342856884002686,0.4937759339809418,1.1716210842132568,1.6919862031936646
+MobileNetV2_fine_tuned,7,0.6249350905418396,0.4958506226539612,1.1754311323165894,1.6897497177124023
+MobileNetV2_fine_tuned,8,0.649350643157959,0.4958506226539612,1.1520062685012817,1.701900601387024
+MobileNetV2_fine_tuned,9,0.6594805121421814,0.49170124530792236,1.1290615797042847,1.6990609169006348
+MobileNetV2_fine_tuned,10,0.6503896117210388,0.4813278019428253,1.1408472061157227,1.6991528272628784

src/config.py

@@ -19,6 +19,39 @@ class Config:
     DROPOUT_RATE = 0.5
     NUM_FROZEN_LAYERS = 2  # Number of layers to freeze in feature extraction
 
+    # Enhanced Architecture Features (set to True to enable)
+    USE_ENHANCED_ARCHITECTURE = False  # Multi-branch architecture with attention
+    USE_PROGRESSIVE_UNFREEZING = False  # Progressive unfreezing strategy
+    USE_ADVANCED_AUGMENTATION = False  # Advanced data augmentation
+    USE_ENSEMBLE_LEARNING = False  # Ensemble learning
+    USE_TEST_TIME_AUGMENTATION = False  # Test-time augmentation
+
+    # Progressive unfreezing parameters
+    PROGRESSIVE_PHASES = 3  # Number of progressive unfreezing phases
+
+    # Ensemble learning parameters
+    ENSEMBLE_STRATEGY = 'soft_voting'  # 'soft_voting', 'hard_voting', 'weighted_voting', 'stacking'
+    ENSEMBLE_MODELS = ["MobileNetV2", "MobileNetV3Large", "MobileNetV3Small"]
+    META_LEARNER_EPOCHS = 50
+
+    # Test-time augmentation parameters
+    TTA_AUGMENTATIONS = 5  # Number of augmentations per test image
+
+    # Advanced training parameters
+    USE_MIXUP = False  # Mixup data augmentation
+    USE_CUTMIX = False  # CutMix data augmentation
+    USE_LABEL_SMOOTHING = False  # Label smoothing
+    LABEL_SMOOTHING_FACTOR = 0.1
+
+    # Learning rate scheduling
+    USE_COSINE_ANNEALING = False  # Cosine annealing scheduler
+    USE_WARM_RESTARTS = False  # Warm restarts
+
+    # Regularization
+    USE_DROPOUT_SCHEDULING = False  # Adaptive dropout scheduling
+    USE_WEIGHT_DECAY = False  # L2 weight decay
+    WEIGHT_DECAY_FACTOR = 1e-4
+
     # Paths
     DATA_DIR = "data"
     MODEL_SAVE_DIR = "saved_models"

src/data_processor.py

@@ -12,12 +12,11 @@ def create_data_generators(self, train_dir, validation_dir=None):
         If validation_dir is not provided, split train_dir using validation_split.
         """
 
-        # Data augmentation pipeline
-        data_augmentation = tf.keras.Sequential([
-            layers.RandomFlip("horizontal"),
-            layers.RandomRotation(0.1),
-            layers.RandomZoom(0.1),
-        ])
+        # Choose augmentation strategy based on config
+        if hasattr(self.config, 'USE_ADVANCED_AUGMENTATION') and self.config.USE_ADVANCED_AUGMENTATION:
+            data_augmentation = self._create_advanced_augmentation()
+        else:
+            data_augmentation = self._create_basic_augmentation()
 
         if validation_dir:
             train_ds = tf.keras.utils.image_dataset_from_directory(
@@ -75,6 +74,62 @@ def create_data_generators(self, train_dir, validation_dir=None):
 
         return train_ds, val_ds
 
+    def _create_basic_augmentation(self):
+        """Create basic data augmentation pipeline."""
+        return tf.keras.Sequential([
+            layers.RandomFlip("horizontal"),
+            layers.RandomRotation(0.1),
+            layers.RandomZoom(0.1),
+        ])
+
+    def _create_advanced_augmentation(self):
+        """Create advanced data augmentation pipeline with more sophisticated techniques."""
+        return tf.keras.Sequential([
+            layers.RandomFlip("horizontal"),
+            layers.RandomRotation(0.2),
+            layers.RandomZoom(0.15),
+            layers.RandomContrast(0.2),
+            layers.RandomBrightness(0.2),
+            layers.RandomTranslation(0.1, 0.1),
+            # Custom augmentation for better leaf/tree feature learning
+            layers.Lambda(lambda x: tf.image.random_hue(x, 0.1)),
+            layers.Lambda(lambda x: tf.image.random_saturation(x, 0.8, 1.2)),
+            layers.Lambda(lambda x: tf.image.random_jpeg_quality(x, 85, 100)),
+        ])
+
+    def create_test_time_augmentation_dataset(self, test_ds, n_augmentations=5):
+        """
+        Create test-time augmentation dataset for improved inference accuracy.
+        """
+        augmentation = tf.keras.Sequential([
+            layers.RandomFlip("horizontal"),
+            layers.RandomRotation(0.05),
+            layers.RandomZoom(0.05),
+            layers.RandomBrightness(0.1),
+        ])
+
+        def augment_batch(x, y):
+            # Create multiple augmented versions of each image
+            augmented_images = []
+            for _ in range(n_augmentations):
+                aug_x = augmentation(x)
+                augmented_images.append(aug_x)
+
+            # Stack all augmented versions
+            stacked = tf.stack(augmented_images, axis=1)  # Shape: (batch, n_aug, height, width, channels)
+
+            # Reshape to treat each augmentation as a separate sample
+            batch_size = tf.shape(x)[0]
+            reshaped = tf.reshape(stacked, (-1, *x.shape[1:]))
+
+            # Repeat labels for each augmentation
+            repeated_labels = tf.repeat(y, n_augmentations, axis=0)
+
+            return reshaped, repeated_labels
+
+        tta_ds = test_ds.map(augment_batch)
+        return tta_ds, n_augmentations
+
     def prepare_dataset_from_directory(self, data_dir):
         """
         Prepare dataset from a directory structure with subdirectories for each class.

src/enhanced_config.py

@@ -0,0 +1,128 @@
+class EnhancedConfig:
+    """
+    Enhanced configuration demonstrating all the new architectural improvements.
+    This configuration enables all advanced features for maximum performance.
+    """
+
+    # Models to train
+    MODELS_TO_TRAIN = ["MobileNetV2"]
+
+    # Data parameters
+    IMG_SIZE = (224, 224)
+    BATCH_SIZE = 32
+    TRAIN_SPLIT = 0.80
+    VALIDATION_SPLIT = 0.25
+
+    # Training parameters
+    INITIAL_EPOCHS = 25
+    FINE_TUNE_EPOCHS = 15
+    INITIAL_LEARNING_RATE = 0.0005
+    FINE_TUNE_LEARNING_RATE = 0.0001
+
+    # Model parameters
+    NUM_CLASSES = 23
+    DROPOUT_RATE = 0.5
+    NUM_FROZEN_LAYERS = 2
+
+    # Enhanced Architecture Features - ALL ENABLED
+    USE_ENHANCED_ARCHITECTURE = True      # Multi-branch architecture with attention
+    USE_PROGRESSIVE_UNFREEZING = True     # Progressive unfreezing strategy
+    USE_ADVANCED_AUGMENTATION = True      # Advanced data augmentation
+    USE_ENSEMBLE_LEARNING = True          # Ensemble learning
+    USE_TEST_TIME_AUGMENTATION = True     # Test-time augmentation
+
+    # Progressive unfreezing parameters
+    PROGRESSIVE_PHASES = 3
+
+    # Ensemble learning parameters
+    ENSEMBLE_STRATEGY = 'stacking'  # Use stacking for best performance
+    ENSEMBLE_MODELS = ["MobileNetV2", "MobileNetV3Large", "MobileNetV3Small"]
+    META_LEARNER_EPOCHS = 50
+
+    # Test-time augmentation parameters
+    TTA_AUGMENTATIONS = 7  # More augmentations for better accuracy
+
+    # Advanced training parameters
+    USE_MIXUP = True
+    USE_CUTMIX = True
+    USE_LABEL_SMOOTHING = True
+    LABEL_SMOOTHING_FACTOR = 0.1
+
+    # Learning rate scheduling
+    USE_COSINE_ANNEALING = True
+    USE_WARM_RESTARTS = True
+
+    # Regularization
+    USE_DROPOUT_SCHEDULING = True
+    USE_WEIGHT_DECAY = True
+    WEIGHT_DECAY_FACTOR = 1e-4
+
+    # Paths
+    DATA_DIR = "data"
+    MODEL_SAVE_DIR = "saved_models"
+    RESULTS_DIR = "results"
+
+
+class ConservativeConfig:
+    """
+    Conservative configuration that enables only the most proven improvements.
+    Good for users who want better performance without experimental features.
+    """
+
+    # Models to train
+    MODELS_TO_TRAIN = ["MobileNetV2"]
+
+    # Data parameters
+    IMG_SIZE = (224, 224)
+    BATCH_SIZE = 32
+    TRAIN_SPLIT = 0.80
+    VALIDATION_SPLIT = 0.25
+
+    # Training parameters
+    INITIAL_EPOCHS = 20
+    FINE_TUNE_EPOCHS = 10
+    INITIAL_LEARNING_RATE = 0.0005
+    FINE_TUNE_LEARNING_RATE = 0.0001
+
+    # Model parameters
+    NUM_CLASSES = 23
+    DROPOUT_RATE = 0.5
+    NUM_FROZEN_LAYERS = 2
+
+    # Enhanced Architecture Features - CONSERVATIVE SELECTION
+    USE_ENHANCED_ARCHITECTURE = False     # Keep original architecture
+    USE_PROGRESSIVE_UNFREEZING = True     # Proven to work well
+    USE_ADVANCED_AUGMENTATION = True      # Almost always beneficial
+    USE_ENSEMBLE_LEARNING = False         # Keep it simple
+    USE_TEST_TIME_AUGMENTATION = True     # Low risk, good reward
+
+    # Progressive unfreezing parameters
+    PROGRESSIVE_PHASES = 3
+
+    # Ensemble learning parameters (not used)
+    ENSEMBLE_STRATEGY = 'soft_voting'
+    ENSEMBLE_MODELS = ["MobileNetV2"]
+    META_LEARNER_EPOCHS = 50
+
+    # Test-time augmentation parameters
+    TTA_AUGMENTATIONS = 5
+
+    # Advanced training parameters
+    USE_MIXUP = False
+    USE_CUTMIX = False
+    USE_LABEL_SMOOTHING = True            # Usually beneficial
+    LABEL_SMOOTHING_FACTOR = 0.1
+
+    # Learning rate scheduling
+    USE_COSINE_ANNEALING = False
+    USE_WARM_RESTARTS = False
+
+    # Regularization
+    USE_DROPOUT_SCHEDULING = False
+    USE_WEIGHT_DECAY = True               # Standard regularization
+    WEIGHT_DECAY_FACTOR = 1e-4
+
+    # Paths
+    DATA_DIR = "data"
+    MODEL_SAVE_DIR = "saved_models"
+    RESULTS_DIR = "results"