Mudassir-08/ModernVGG-BN-CIFAR10

---

ModernVGG-BN CIFAR-10 Image Classification (PyTorch)

---

Project Overview

This project implements a VGG-style Convolutional Neural Network for image classification on CIFAR-10 using PyTorch.

The model is trained from scratch with modern deep learning techniques including:

• Mixed Precision Training (AMP)
• AdamW optimizer
• Cosine Annealing Learning Rate Scheduler
• Gradient Clipping
• Early Stopping
• Batch Normalization throughout the network

---

Dataset

• CIFAR-10 dataset
• 10 classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck
• 32×32 RGB images

---

Model Architecture

Feature Extractor (VGG-style CNN)

64 → 128 → 256 → 512 → 512

Each block: Conv2D → BatchNorm → ReLU → Conv2D → BatchNorm → ReLU → MaxPool

---

Classifier Head

• AdaptiveAvgPool2D (1×1)
• Flatten
• Linear (512 → 256)
• BatchNorm1D
• ReLU
• Dropout (0.2)
• Linear (256 → 10)

---

Training Pipeline

Each epoch includes:

1. Forward pass
2. Loss computation
3. Backpropagation
4. Gradient clipping
5. Optimizer step (AdamW)
6. Scheduler step (CosineAnnealingLR)
7. Validation
8. Metric logging
9. Best model checkpoint saving
10. Early stopping monitoring

---

Regularization Techniques

• Batch Normalization
• Dropout (0.2)
• Gradient Clipping
• Early Stopping

---

Results

• CIFAR-10.1 Accuracy: 78.95%
• Strong generalization on unseen dataset (CIFAR-10.1)
• Most confusion between:
  • automobile ↔ truck
  • cat ↔ dog

---

Inference Pipeline

Image → Resize → Normalize → Model → Softmax → Prediction

Outputs:

• Predicted class
• Confidence score
• Top-5 probabilities

---

Model Saving Format

torch.save({
    "epoch": epoch,
    "model_state_dict": model.state_dict(),
    "optimizer_state_dict": optimizer.state_dict(),
    "scheduler_state_dict": scheduler.state_dict(),
    "scaler_state_dict": scaler.state_dict(),
    "best_acc": best_acc
}, "best_model.pth")