PAVIC Augmentation Research Pipeline

Overview

Research pipeline for evaluating synthetic data generation via fine-tuned diffusion models on class-imbalanced image classification. Compares ResNet and Vision Transformer (ViT) performance across three scenarios:

1. Imbalanced datasets (real data only)
2. Traditional oversampling (duplicated minority samples)
3. Synthetic augmentation (real + diffusion-generated minority samples)

Research Pipeline

9-stage experimental pipeline:

1. Dataset Preparation: AdImageNet, CIFAR-FS, and plant pathology datasets
2. Imbalance Induction: Systematic minority class sample removal
3. Diffusion Fine-tuning: Fine-tune models on minority samples
4. Synthetic Generation: Generate minority samples to match majority counts
5. Dataset Construction: Create three dataset versions
6. Classifier Training: Train ResNet and ViT on each version
7. Evaluation: Comprehensive metrics and confusion matrices
8. Analysis: Performance comparison and quality assessment
9. Documentation: Reproducible results with detailed logging

Installation

Requirements

• Python 3.8+, GPU (recommended, auto-detects CPU fallback), 8GB+ RAM, 20GB+ disk space

Setup

git clone <repository-url>
cd pavic-augmentation
pip install -r requirements.txt

Usage

Quick Start

python main.py

Options

# Specific datasets
python main.py --datasets ad-imagenet cifar-fs

# Skip stages for testing
python main.py --skip-diffusion
python main.py --skip-training --skip-evaluation

# Custom experiment
python main.py --experiment-name "my_experiment" --seed 123

Command Line Options

• --experiment-name: Custom experiment name
• --datasets: Choose datasets (ad-imagenet, cifar-fs, plant-pathology)
• --skip-diffusion: Skip diffusion fine-tuning and synthetic generation
• --skip-training: Skip classifier training
• --skip-evaluation: Skip model evaluation
• --seed: Random seed (default: 42)

Configuration

Edit src/config.py for key parameters:

• datasets: Dataset selection
• minority_classes: Classes to make minority
• imbalance_ratio: Fraction of minority samples to keep
• diffusion_model: Base diffusion model
• num_epochs: Training epochs
• learning_rate: Learning rate

Datasets

1. AdImageNet: 9,003 advertisement images (auto-downloaded)
2. CIFAR-FS: Place FC100 pickle files in data/raw/cifar-fs/
3. Plant Pathology: Generated locally for testing

Results

Generated in results/experiment_name/:

• Performance metrics (accuracy, F1, confusion matrices)
• Comparison visualizations and tables
• Class distribution analysis
• Synthetic image quality assessment

Hardware

• RTX 4070: Automatically detected and optimized (12GB VRAM)
• Other GPUs: 8GB+ VRAM recommended (auto-detects)
• CPU: Automatically adjusts batch size and epochs

GPU Setup

# Quick check if your RTX 4070 is detected
python quick_gpu_check.py

# Full GPU setup (if needed)
python check_gpu.py

Troubleshooting

• GPU not detected: Run python check_gpu.py to install CUDA PyTorch
• Import errors: pip install --upgrade diffusers transformers accelerate
• Memory issues: Automatically optimized per hardware
• Testing: Use --skip-diffusion for faster runs