unlearning LF

Data.py

@@ -1,8 +1,10 @@
 from torchvision import datasets, transforms, models
+from torch.utils.data import Dataset, DataLoader, Subset
 import torch
 import numpy as np
 
 
+
 # train set transform 
 def train_transform(res):
     return transforms.Compose([
@@ -101,3 +103,39 @@ def get_indices(dataset, identities, split_at, size = 30):
         test_indices.extend(indices[split_at:size])
 
     return train_indices, test_indices
+
+
+
+def get_forget_retain_loaders(dataset: Dataset, forget_class_idx: int, batch_size: int = 32) -> tuple[DataLoader, DataLoader]:
+    """
+    Splits an IdentitySubset or standard Dataset into forget and retain sets 
+    based on a remapped target class index.
+    """
+    # 1. Safely extract targets whether it's a standard dataset or a Subset wrapper
+    if hasattr(dataset, 'targets'):
+        targets = dataset.targets
+    elif hasattr(dataset, 'identity'):  # Raw CelebA support
+        targets = dataset.identity
+    else:
+        # If it's an IdentitySubset or standard Subset, extract mapped targets sequentially
+        # This guarantees we get the 0 -> (n-1) remapped labels
+        targets = [dataset[i][1] for i in range(len(dataset))]
+        
+    if not isinstance(targets, torch.Tensor):
+        targets = torch.tensor(targets)
+        
+    # 2. Generate mask indices local to this subset
+    forget_indices = torch.where(targets == forget_class_idx)[0].tolist()
+    retain_indices = torch.where(targets != forget_class_idx)[0].tolist()
+    
+    # 3. Create PyTorch Subsets
+    forget_subset = Subset(dataset, forget_indices)
+    retain_subset = Subset(dataset, retain_indices)
+    
+    # 4. Wrap into clean DataLoaders
+    forget_loader = DataLoader(forget_subset, batch_size=batch_size, shuffle=False)
+    retain_loader = DataLoader(retain_subset, batch_size=batch_size, shuffle=True)
+    
+    print(f"[Data Split] Local Class {forget_class_idx}: {len(forget_subset)} samples | Remaining Classes: {len(retain_subset)} samples.")
+    
+    return forget_loader, retain_loader

LinearFiltration_metrics.txt

@@ -0,0 +1,7 @@
+execution_time_sec
+0.000996
+0.030071
+0.001182
+0.001176
+0.001229
+0.001257

Tune.py

@@ -7,12 +7,13 @@ from sklearn.metrics import classification_report
 import SetUp
 from Data import *
 #from datasets.Casia import *
-#from IdentitySubset import IdentitySubset
-from datasets.UniversalIdentitySubset import UniversalIdentitySubset as IdentitySubset
+from IdentitySubset import IdentitySubset
+#from datasets.UniversalIdentitySubset import UniversalIdentitySubset as IdentitySubset
 # models
 from architectures.Model import Model, Architecture
 
-from unlearning import LinearFiltration, WeightFiltration, CertifiedRemoval
+from unlearning.LinearFiltration import LinearFiltration
+# WeightFiltration, CertifiedRemoval
 
 # numbre of classes 
 CLASS_SIZE = 20
@@ -24,7 +25,7 @@ BATCH_SIZE = 32
 SAMPLE_SIZE = 30
 
 # this is then (full_sample - test_sample)
-TRAINING_SMPLE = 28
+TRAINING_SMPLE = 27
 
 # learning rate
 LR_RATE = 0.0001
@@ -96,68 +97,79 @@ print(f'> Constants : Classes = {CLASS_SIZE}, Batch = {BATCH_SIZE}, epochs = {EP
 # MODEL PREPARATION
 # cuda if exists (it does here)
 device = SetUp.get_device()
+
+for i in range(0,CLASS_SIZE):
 # Create model using Factory
-model = Model.create(
-    arch = arch, 
-    device = device, 
-    size = CLASS_SIZE)
+    model = Model.create(
+        arch = arch, 
+        device = device, 
+        size = CLASS_SIZE)
 
-# we may need to load existing model or finetune
-model.train(
-        epochs = EPOCHS, 
-        loader = train_loader, 
-        rate = LR_RATE)
+    # we may need to load existing model or finetune
+    model.train(
+            epochs = EPOCHS, 
+            loader = train_loader, 
+            rate = LR_RATE)
 
-    # save.
-model.save(filename=arch.name.lower())
+        # save.
+    model.save(filename=arch.name.lower())
 
 
-# done tuning
-print('Model saved!')
+    # done tuning
 
-# EVALUATE
+    # EVALUATE
+    te_transform = test_transform(RESOLUTION)
+    # Testing
+    test_data = IdentitySubset(
+        dataset = dataset, 
+        indices=test_indices, 
+        id_mapping=id_map,
+        transform=te_transform)
 
-te_transform = test_transform(RESOLUTION)
-# Testing
-test_data = IdentitySubset(
-    dataset = dataset, 
-    indices=test_indices, 
-    id_mapping=id_map,
-    transform=te_transform)
+    test_loader = DataLoader(
+        test_data, 
+        batch_size=BATCH_SIZE, 
+        shuffle=False)
 
-test_loader = DataLoader(
-    test_data, 
-    batch_size=BATCH_SIZE, 
-    shuffle=False)
+    print(f"Total test images for these {CLASS_SIZE} classes: {len(test_data)}")
 
-print(f"Total test images for these {CLASS_SIZE} classes: {len(test_data)}")
+    # Evaluate
+    model.evaluate(
+        loader = test_loader,
+        mode="finetunned"
+        )
 
-# Evaluate
-model.evaluate(
-    loader = test_loader)
+    # test again
+    reloaded = Model.create(
+        arch=arch,
+        device = device,
+        size = CLASS_SIZE
+        )
+    reloaded.load(arch = arch)
+    print("fine tunned model loaded")
+    # reloaded.evaluate(
+    #   loader = test_loader
+    #)
 
-# test again
-reloaded = Model.create(
-    arch=arch,
-    device = device,
-    size = CLASS_SIZE
+    # Unlearning
+    FORGET_CLASS_IDX = i
+
+    forget_test_loader, retain_test_loader = get_forget_retain_loaders(
+        dataset=test_data, 
+        forget_class_idx=FORGET_CLASS_IDX, 
+        batch_size=BATCH_SIZE
     )
-reloaded.load(arch = arch)
-print("Evaluating loaded")
-reloaded.evaluate(
-    loader = test_loader
-)
+
+    #retain_test_loader = DataLoader(retain_test_loader.dataset, batch_size=BATCH_SIZE, shuffle=False)
+
+    # 3. Instantiate and apply the Linear Filtration rule
+    filtration = LinearFiltration(target_class_idx=FORGET_CLASS_IDX)
+    filtration.apply(reloaded.model)
 
 
-strategies_to_test = [
-    LinearFiltration(target_class_idx=12),
-    WeightFiltration(target_class_idx=12),
-    CertifiedRemoval(target_class_idx=12)
-]
+    # 4. Final Performance Analysis
+    print("\n--- Performance on Retained Classes")
+    reloaded.evaluate(loader=retain_test_loader, mode="retain")
 
-# Run the comparative benchmark seamlessly
-execution_profiles = {}
-for strategy in strategies_to_test:
-    # Each iteration clones weights back to fine-tuned state before running
-    runtime = my_model.unlearn(strategy, forget_loader, retain_loader)
-    execution_profiles[strategy.__class__.__name__] = runtime
+    print("\n--- Performance on Forgotten Class")
+    reloaded.evaluate(loader=forget_test_loader,mode="forget")

architectures/Model.py

@@ -43,10 +43,13 @@ class Model(ABC):
         if self.device.type == 'cuda': torch.cuda.synchronize()
         print(f"Training completed in: {time.time() - start_time:.2f}s")
 
+
     def evaluate(self, loader):
         self.model.eval()
         all_preds, all_labels = [], []
         print("\nEvaluating...")
+
+        
         
         with torch.no_grad():
             for inputs, labels in loader:
@@ -56,9 +59,11 @@ class Model(ABC):
                 all_preds.extend(predicted.cpu().numpy())
                 all_labels.extend(labels.cpu().numpy())
 
-        accuracy = 100 * (np.array(all_preds) == np.array(all_labels)).sum() / len(all_labels)
+        classes = sorted(list(set(all_labels)))
+
+        accuracy = 100 * (np.array(all_preds) == np.array(all_labels)).sum() / len(classes)
         print(f"Test Accuracy: {accuracy:.2f}%")
-        print(classification_report(all_labels, all_preds, zero_division=0))
+        print(classification_report(all_labels, all_preds, labels=classes, zero_division=0))
 
 
     def save(self, filename=None):
@@ -101,13 +106,109 @@ class Model(ABC):
         start_time = time.time()
         
         # Delegate the actual algorithmic weight/logit manipulation to the strategy
-        strategy.apply(self.network, forget_loader, retain_loader)
+        strategy.apply(self.model, forget_loader, retain_loader)
         
         elapsed_time = time.time() - start_time
         print(f"{strategy.__class__.__name__} completed in {elapsed_time:.4f} seconds.")
         
         return elapsed_time
 
+    def evaluate(self, loader, mode="eval"):
+        """
+        Evaluates the model, prints terminal reports, and routes metrics to 
+        a file logger based on the current context mode.
+        """
+        self.model.eval()
+        all_preds, all_labels = [], []
+        print(f"\nEvaluating Domain: [{mode}]...")
+        
+        with torch.no_grad():
+            for inputs, labels in loader:
+                inputs, labels = inputs.to(self.device), labels.to(self.device)
+                outputs = self.model(inputs)
+                _, predicted = torch.max(outputs, 1)
+                all_preds.extend(predicted.cpu().numpy())
+                all_labels.extend(labels.cpu().numpy())
+
+        # Extract only the active classes evaluated in this loader slice
+        classes = sorted(list(set(all_labels)))
+        accuracy = 100 * (np.array(all_preds) == np.array(all_labels)).sum() / len(all_labels)
+        
+        print(f"Test Accuracy: {accuracy:.2f}%")
+        
+        # 1. Print standard text report to terminal
+        print(classification_report(all_labels, all_preds, labels=classes, zero_division=0))
+        
+        # 2. Extract structured dictionary metrics
+        report_dict = classification_report(
+            all_labels, 
+            all_preds, 
+            labels=classes, 
+            output_dict=True, 
+            zero_division=0
+        )
+        
+        # 3. Delegate file tracking to isolated helper method
+        self._log_to_csv(mode, accuracy, classes, report_dict)
+
+
+    def _log_to_csv(self, mode, accuracy, classes, report_dict):
+        """Handles directory structures, file setups, and distinct CSV column formatting."""
+        arch_name = self.__class__.__name__.lower()
+        save_dir = Path("reports")
+        save_dir.mkdir(parents=True, exist_ok=True)
+        csv_path = save_dir / f"{arch_name}-{mode}.csv"
+        
+        file_exists = csv_path.exists()
+        '''
+        # Structure payload and headers based on evaluation slice type
+        if mode == "forget":
+            headers = ["accuracy", "precision", "recall", "f1-score"]
+            target_cls_str = str(classes[0])
+            metrics = report_dict[target_cls_str]
+            row = [
+                f"{accuracy / 100.0:.4f}",
+                f"{metrics['precision']:.4f}",
+                f"{metrics['recall']:.4f}",
+                f"{metrics['f1-score']:.4f}"
+            ]
+        else:
+            headers = [
+                "accuracy", 
+                "macro_precision", "macro_recall", "macro_f1",
+                "weighted_precision", "weighted_recall", "weighted_f1"
+            ]
+            row = [
+                f"{accuracy / 100.0:.4f}",
+                f"{report_dict['macro avg']['precision']:.4f}",
+                f"{report_dict['macro avg']['recall']:.4f}",
+                f"{report_dict['macro avg']['f1-score']:.4f}",
+                f"{report_dict['weighted avg']['precision']:.4f}",
+                f"{report_dict['weighted avg']['recall']:.4f}",
+                f"{report_dict['weighted avg']['f1-score']:.4f}"
+            ]'''
+        headers = [
+                "accuracy", 
+                "macro_precision", "macro_recall", "macro_f1",
+                "weighted_precision", "weighted_recall", "weighted_f1"
+        ]
+        row = [
+            f"{accuracy / 100.0:.4f}",
+            f"{report_dict['macro avg']['precision']:.4f}",
+            f"{report_dict['macro avg']['recall']:.4f}",
+            f"{report_dict['macro avg']['f1-score']:.4f}",
+            f"{report_dict['weighted avg']['precision']:.4f}",
+            f"{report_dict['weighted avg']['recall']:.4f}",
+            f"{report_dict['weighted avg']['f1-score']:.4f}"
+        ]
+            
+        with open(csv_path, "a") as f:
+            if not file_exists:
+                f.write(",".join(headers) + "\n")
+            f.write(",".join(row) + "\n")
+            
+        print(f">> Direct CSV metrics appended to {csv_path}")
+
 
 
     # Using the factory patern here

reports/LinearFiltration_metrics.txt

@@ -0,0 +1,35 @@
+execution_time_sec
+0.001269
+0.001227
+0.001298
+0.001281
+0.001178
+0.001392
+0.001391
+0.001338
+0.001162
+0.001355
+0.001361
+0.001241
+0.001210
+0.001152
+0.001358
+0.001250
+0.001467
+0.001248
+0.001411
+0.001470
+0.001241
+0.001366
+0.001206
+0.001339
+0.001268
+0.002847
+0.001245
+0.001299
+0.001222
+0.001274
+0.001351
+0.001401
+0.001286
+0.001214

reports/resnet50-finetunned.csv

@@ -0,0 +1,21 @@
+accuracy,macro_precision,macro_recall,macro_f1,weighted_precision,weighted_recall,weighted_f1
+0.9333,0.9500,0.9333,0.9264,0.9500,0.9333,0.9264
+0.9167,0.9425,0.9167,0.9111,0.9425,0.9167,0.9111
+0.9333,0.9500,0.9333,0.9264,0.9500,0.9333,0.9264
+0.8667,0.9050,0.8667,0.8625,0.9050,0.8667,0.8625
+0.9000,0.9208,0.9000,0.8976,0.9208,0.9000,0.8976
+0.9000,0.9208,0.9000,0.8926,0.9208,0.9000,0.8926
+0.9667,0.9750,0.9667,0.9657,0.9750,0.9667,0.9657
+0.9000,0.9308,0.9000,0.9012,0.9308,0.9000,0.9012
+0.9833,0.9875,0.9833,0.9829,0.9875,0.9833,0.9829
+0.9500,0.9625,0.9500,0.9486,0.9625,0.9500,0.9486
+0.8667,0.9008,0.8667,0.8551,0.9008,0.8667,0.8551
+0.9167,0.9375,0.9167,0.9093,0.9375,0.9167,0.9093
+0.9000,0.9250,0.9000,0.8921,0.9250,0.9000,0.8921
+0.9000,0.9333,0.9000,0.9024,0.9333,0.9000,0.9024
+0.9500,0.9625,0.9500,0.9486,0.9625,0.9500,0.9486
+0.9167,0.9333,0.9167,0.9148,0.9333,0.9167,0.9148
+0.9167,0.9375,0.9167,0.9143,0.9375,0.9167,0.9143
+0.9667,0.9750,0.9667,0.9657,0.9750,0.9667,0.9657
+0.9333,0.9500,0.9333,0.9314,0.9500,0.9333,0.9314
+0.9000,0.9350,0.9000,0.8957,0.9350,0.9000,0.8957

reports/resnet50-forget.csv

@@ -0,0 +1,21 @@
+accuracy,macro_precision,macro_recall,macro_f1,weighted_precision,weighted_recall,weighted_f1
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000
+0.0000,0.0000,0.0000,0.0000,0.0000,0.0000,0.0000

reports/resnet50-retain.csv

@@ -0,0 +1,21 @@
+accuracy,macro_precision,macro_recall,macro_f1,weighted_precision,weighted_recall,weighted_f1,
+0.9474,0.9605,0.9474,0.9459,0.9605,0.9474,0.9459
+0.9123,0.9395,0.9123,0.9064,0.9395,0.9123,0.9064
+0.9298,0.9526,0.9298,0.9244,0.9526,0.9298,0.9244
+0.8596,0.9000,0.8596,0.8553,0.9000,0.8596,0.8553
+0.9123,0.9298,0.9123,0.9103,0.9298,0.9123,0.9103
+0.9123,0.9342,0.9123,0.9045,0.9342,0.9123,0.9045
+0.9649,0.9737,0.9649,0.9639,0.9737,0.9649,0.9639
+0.8947,0.9272,0.8947,0.8960,0.9272,0.8947,0.8960
+1.0000,1.0000,1.0000,1.0000,1.0000,1.0000,1.0000
+0.9649,0.9737,0.9649,0.9639,0.9737,0.9649,0.9639
+0.8772,0.9132,0.8772,0.8650,0.9132,0.8772,0.8650
+0.9123,0.9342,0.9123,0.9045,0.9342,0.9123,0.9045
+0.9123,0.9342,0.9123,0.9045,0.9342,0.9123,0.9045
+0.9298,0.9605,0.9298,0.9328,0.9605,0.9298,0.9328
+0.9649,0.9737,0.9649,0.9639,0.9737,0.9649,0.9639
+0.9298,0.9430,0.9298,0.9283,0.9430,0.9298,0.9283
+0.9298,0.9474,0.9298,0.9278,0.9474,0.9298,0.9278
+0.9649,0.9737,0.9649,0.9639,0.9737,0.9649,0.9639
+0.9298,0.9474,0.9298,0.9278,0.9474,0.9298,0.9278
+0.8947,0.9316,0.8947,0.8902,0.9316,0.8947,0.8902

unlearning/LinearFiltration.py

@@ -1,29 +1,48 @@
 
 import torch
-from Strategy import Strategy
+import torch.nn as nn
+from .Strategy import Strategy
 
-class NormalizingLinearFiltration(Strategy):
-    def __init__(self, target_class_idx):
+class LinearFiltration(Strategy):
+    def __init__(self, target_class_idx: int):
+        super().__init__()  # Automatically configures 'NormalizingLinearFiltration_metrics.txt'
         self.target_class_idx = target_class_idx
 
-    def apply(self, model, forget_loader, retain_loader):
+    def _run(self, model: nn.Module) -> nn.Module:
         model.eval()
-        # Freeze parameters structurally
         for param in model.parameters():
             param.requires_grad = False
             
         with torch.no_grad():
-            # we modify only classification head
-            # Shape: [num_classes, feature_dim]
-            W = model.fc.weight.data
-            
-            # Compute the normalization transformation projection matrix (A)
-            # (In your full code, calculate A here matching Baumhauer et al.'s equations)
+            W = model.fc.weight.data.clone()
             num_classes = W.shape[0]
-            A = torch.eye(num_classes, device=W.device)
-            # Mask/blend target class index distribution configurations here...
-            A[self.target_class_idx, :] = 0.0 
             
-            # 3. Direct weight matrix override: W_filtered = A * W
+            A = self._calculate_filtration_matrix(num_classes, self.target_class_idx, W.device)
             sanitized_W = torch.mm(A, W)
-            model.fc.weight.copy_(sanitized_W)
+            model.fc.weight.copy_(sanitized_W)
+            
+        return model
+
+    '''@staticmethod
+    def _calculate_filtration_matrix(num_classes: int, forget_class: int, device: torch.device) -> torch.Tensor:
+        A = torch.eye(num_classes, device=device)
+        num_remaining_classes = num_classes - 1
+        for j in range(num_classes):
+            if j == forget_class:
+                A[forget_class, j] = 0.0
+            else:
+                A[forget_class, j] = 1.0 / num_remaining_classes
+        return A'''
+    @staticmethod
+    def _calculate_filtration_matrix(num_classes: int, forget_class: int, device: torch.device) -> torch.Tensor:
+        A = torch.eye(num_classes, device=device)
+        num_remaining = num_classes - 1
+        
+        # The row of the forgotten class should average the inputs of all other classes
+        for j in range(num_classes):
+            if j == forget_class:
+                A[forget_class, j] = 0.0
+            else:
+                A[forget_class, j] = 1.0 / num_remaining
+                
+        return A

unlearning/Strategy.py

@@ -0,0 +1,47 @@
+
+import torch.nn as nn
+import time
+import os
+
+class Strategy:
+    """Abstract base class for unlearning algorithms with automated, strategy-specific logging."""
+    
+    def __init__(self):
+        # Dynamically set file name based on the class name (e.g., 'NormalizingLinearFiltration.txt')
+        self.strategy_name = self.__class__.__name__
+        self.log_file = f"reports/{self.strategy_name}_metrics.txt"
+        self._initialize_log_file()
+
+    def _initialize_log_file(self):
+        """Creates a unique log file for this strategy with a header if it doesn't exist."""
+        if not os.path.exists(self.log_file):
+            with open(self.log_file, "w") as f:
+                f.write("execution_time_sec\n")
+
+    def log_metric(self, execution_time: float):
+        """Appends the execution time to this strategy's specific file."""
+        with open(self.log_file, "a") as f:
+            f.write(f"{execution_time:.6f}\n")
+
+    def apply(self, model: nn.Module) -> nn.Module:
+        """
+        Wraps the unlearning execution with automated timing and strategy-specific logging.
+        DO NOT override this method in subclasses. Override _run instead.
+        """
+        start_time = time.perf_counter()
+        
+        # Execute core unlearning logic
+        processed_model = self._run(model)
+        
+        end_time = time.perf_counter()
+        execution_time = end_time - start_time
+        
+        # Log to the strategy's specific file
+        self.log_metric(execution_time)
+        print(f"[{self.strategy_name}] Completed in {execution_time:.6f} seconds. Saved to {self.log_file}")
+        
+        return processed_model
+
+    def _run(self, model: nn.Module) -> nn.Module:
+        """Subclasses implement their core unlearning logic here."""
+        raise NotImplementedError