CIFAR-10 Object Detection Using CNN

Project Overview

This project implements a Convolutional Neural Network (CNN) for object detection on the CIFAR-10 dataset using PyTorch and transfer learning with MobileNetV3. The model achieves high accuracy in classifying images into 10 different categories through a two-stage training approach with feature freezing and fine-tuning.

Dataset

CIFAR-10 Dataset contains 60,000 32x32 color images in 10 classes:

• Airplane
• Automobile
• Bird
• Cat
• Deer
• Dog
• Frog
• Horse
• Ship
• Truck

Dataset Split:

• Training: 50,000 images
• Testing: 10,000 images

Architecture

Base Model: MobileNetV3-Large

• Pre-trained weights: ImageNet weights for transfer learning
• Modified classifier: Custom classifier head for 10 classes
• Input size: 224x224 pixels (resized from 32x32)

Model Architecture Details:

MobileNetV3(
  features: Pre-trained feature extractor (960 output features)
  classifier: Sequential(
    Linear(960 → 1280)
    Hardswish()
    Dropout(0.2)
    Linear(1280 → 10)  # 10 CIFAR-10 classes
  )
)

Implementation Details

Data Preprocessing

• Image resizing: 32x32 → 224x224 for MobileNet compatibility
• Normalization: ImageNet standards (mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225])
• Data format: PNG images extracted from 7z archives

Training Strategy

1. Stage 1 - Feature Extraction (15 epochs)

   • Freeze feature extractor layers
   • Train only classifier head
   • Learning rate: 0.001

2. Stage 2 - Fine-tuning (15 epochs)

   • Unfreeze all layers
   • End-to-end training
   • Learning rate: 0.001

Training Configuration

• Optimizer: Adam
• Loss function: CrossEntropyLoss
• Batch size: 64
• Total epochs: 30 (15 + 15)
• Device: CUDA-enabled GPU

Performance Metrics

The model tracks:

• Training Loss: Monitored every 100 batches
• Training Accuracy: Real-time accuracy calculation
• Running metrics: Loss and accuracy averages

Training Progress Monitoring

def train_one_epoch(dataloader, model, loss_fn, optimizer):
    # Real-time tracking of:
    # - Batch-wise loss
    # - Running accuracy
    # - Progress indicators

Key Features

1. Transfer Learning

• Leverages pre-trained MobileNetV3 for efficient training
• Reduces training time and improves convergence

2. Progressive Training

• Two-stage approach prevents overfitting
• Gradual unfreezing for better feature adaptation

3. Custom Dataset Classes

class TrainDataset(Dataset):
    # Handles image loading, preprocessing, and label encoding

class TestDataset(Dataset):
    # Processes test images for inference

4. Efficient Data Pipeline

• Automated 7z archive extraction
• On-the-fly image preprocessing
• Optimized DataLoader configuration

Project Structure

CIFAR-10-Object-Detection-Using-CNN/
├── CIFAR-10-Object-Detection-Using-CNN.ipynb
├── .gitignore
└── README.md

Usage

Prerequisites

pip install torch torchvision py7zr pandas numpy

Running the Model

1. Data Setup: Place CIFAR-10 datasets in specified paths
2. Training: Execute the notebook cells sequentially
3. Evaluation: Model automatically generates predictions on test set

Key Code Components

Model Definition

class Cifar10(nn.Module):
    def __init__(self):
        super().__init__()
        self.pretrainednet = mobilenet_v3_large(weights=MobileNet_V3_Large_Weights.DEFAULT)
        # Custom classifier for 10 classes

Training Loop

for i in range(n_epochs):
    train_epoch_loss, train_epoch_acc = train_one_epoch(
        traindataloader, model, loss_fn, optimizer
    )

Technical Highlights

1. Memory Efficient Architecture

• MobileNetV3 designed for mobile/edge deployment
• Optimized depth-wise separable convolutions

2. Robust Preprocessing

• Handles various image formats
• Automatic normalization and resizing

3. Flexible Training Pipeline

• Configurable hyperparameters
• Easy modification for different datasets

Results

The model produces:

• Submission file: CSV with predicted class labels
• Performance metrics: Training loss and accuracy curves
• Class predictions: Mapped back to original CIFAR-10 labels

Future Improvements

1. Data Augmentation: Add rotation, flip, and color jittering
2. Advanced Architectures: Experiment with ResNet, EfficientNet
3. Ensemble Methods: Combine multiple models for better accuracy
4. Learning Rate Scheduling: Implement adaptive learning rates
5. Cross-validation: Add k-fold validation for robust evaluation

Dependencies

• PyTorch: Deep learning framework
• torchvision: Computer vision utilities and models
• py7zr: 7z archive extraction
• pandas: Data manipulation
• numpy: Numerical computations

Key Achievements

• Successful implementation of transfer learning
• Two-stage training strategy
• Efficient data pipeline for large datasets
• GPU-accelerated training
• Production-ready inference pipeline

---