added material

privacy/c1.1.py

@@ -0,0 +1,127 @@
+#!/usr/bin/env python3
+
+import os
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.optim as optim
+from torch.utils.data import DataLoader
+from torchvision import datasets, transforms
+from tqdm import tqdm
+from opacus import PrivacyEngine
+from opacus.validators import ModuleValidator
+
+# =============================================================================
+# STABLE CONFIGURATION
+# =============================================================================
+BATCH_SIZE = 64      # Smaller batch = more stable memory on CPU
+DP_EPOCHS = 10       # Fewer epochs to start; check stability first
+DP_LR = 5e-4         # Lower LR to prevent 'nan'
+MAX_GRAD_NORM = 1.0
+DELTA = 1e-5
+TARGET_EPSILON = 6.0
+DEVICE = torch.device('cpu')
+
+# =============================================================================
+# MODEL (MINIMAL & STABLE)
+# =============================================================================
+class SimpleSVHN(nn.Module):
+    def __init__(self):
+        super().__init__()
+        # We use GroupNorm with small groups for maximum stability
+        self.conv1 = nn.Conv2d(3, 16, 3, padding=1)
+        self.gn1 = nn.GroupNorm(2, 16)
+        self.conv2 = nn.Conv2d(16, 32, 3, padding=1)
+        self.gn2 = nn.GroupNorm(4, 32)
+        self.pool = nn.MaxPool2d(2, 2)
+        self.fc = nn.Linear(32 * 8 * 8, 10)
+
+    def forward(self, x):
+        x = self.pool(F.relu(self.gn1(self.conv1(x)))) # 16x16
+        x = self.pool(F.relu(self.gn2(self.conv2(x)))) # 8x8
+        x = torch.flatten(x, 1)
+        return self.fc(x)
+
+# =============================================================================
+# TRAINING SCRIPT
+# =============================================================================
+def run_stable_dp_training():
+    # 1. Data Loading
+    transform = transforms.Compose([
+        transforms.ToTensor(),
+        transforms.Normalize((0.4377, 0.4438, 0.4728), (0.1980, 0.2010, 0.1970))
+    ])
+    
+    train_ds = datasets.SVHN("data", split='train', download=True, transform=transform)
+    test_ds = datasets.SVHN("data", split='test', download=True, transform=transform)
+    
+    train_loader = DataLoader(train_ds, batch_size=BATCH_SIZE, shuffle=True)
+    test_loader = DataLoader(test_ds, batch_size=BATCH_SIZE, shuffle=False)
+
+    # 2. Model Setup
+    model = SimpleSVHN().to(DEVICE)
+    model = ModuleValidator.fix(model)
+    
+    optimizer = optim.Adam(model.parameters(), lr=DP_LR)
+
+    # 3. Privacy Engine (THE STABILITY FIX)
+    # Using grad_sample_mode="hooks" prevents the C++ level SegFaults
+    privacy_engine = PrivacyEngine()
+    model, optimizer, train_loader = privacy_engine.make_private_with_epsilon(
+        module=model,
+        optimizer=optimizer,
+        data_loader=train_loader,
+        target_epsilon=TARGET_EPSILON,
+        target_delta=DELTA,
+        epochs=DP_EPOCHS,
+        max_grad_norm=MAX_GRAD_NORM,
+        grad_sample_mode="hooks", 
+    )
+
+    criterion = nn.CrossEntropyLoss()
+
+    print(f"Starting Training on {DEVICE}...")
+    
+    for epoch in range(DP_EPOCHS):
+        model.train()
+        total_loss = 0
+        
+        pbar = tqdm(train_loader, desc=f"Epoch {epoch+1}")
+        for images, labels in pbar:
+            optimizer.zero_grad()
+            
+            outputs = model(images)
+            loss = criterion(outputs, labels)
+            
+            if torch.isnan(loss):
+                print("\n[!] Warning: NaN detected. Skipping batch.")
+                continue
+                
+            loss.backward()
+            optimizer.step()
+            
+            total_loss += loss.item()
+            pbar.set_postfix({'loss': f'{loss.item():.2f}'})
+
+        epsilon = privacy_engine.get_epsilon(DELTA)
+        print(f"Epoch {epoch+1} complete. Current ε: {epsilon:.2f}")
+
+    # 4. Final Evaluation
+    model.eval()
+    correct = 0
+    total = 0
+    with torch.no_grad():
+        for images, labels in test_loader:
+            images, labels = images.to(DEVICE), labels.to(DEVICE)
+            outputs = model(images)
+            _, predicted = outputs.max(1)
+            total += labels.size(0)
+            correct += predicted.eq(labels).sum().item()
+
+    print(f"\nFinal Test Accuracy: {100.*correct/total:.2f}%")
+    print(f"Final Epsilon: {epsilon:.2f}")
+
+if __name__ == "__main__":
+    # Force single-thread for math if your CPU is still fighting the memory allocation
+    torch.set_num_threads(1) 
+    run_stable_dp_training()