leekwoon/diffig_gig_backup

DiffIG GIG Backup

Research project backup for DiffIG (Diffusion-based Integrated Gradients) and GIG (Guided Integrated Gradients) - advanced explainability methods for image classifiers.

Download & Extract

# huggingface_hub 설치 (필요시)
pip install huggingface_hub

# 다운로드
huggingface-cli download leekwoon/diffig_gig_backup --repo-type dataset --local-dir ./diffig_data

# 무결성 확인 (선택사항)
cd diffig_data
md5sum -c checksums.md5

# 파일 합치기 및 압축 해제
cat data.tar.gz.part_* | tar -xzvf -

Directory Structure

diffig/
├── diffig/                 # Core implementation modules
│   ├── explainer/          # Explainability algorithms
│   │   ├── ig.py           # Integrated Gradients
│   │   ├── ig2.py          # Improved IG
│   │   ├── gig.py          # Guided IG
│   │   ├── agi.py          # Adversarial GI
│   │   ├── diffig.py       # DiffIG (our method)
│   │   ├── mig.py          # Manifold IG
│   │   ├── big.py          # Boundary IG
│   │   ├── eig.py          # Enhanced IG
│   │   ├── sg.py           # Smooth Gradients
│   │   ├── spi.py          # SmoothGrad Path Integration
│   │   ├── fullgrad.py     # FullGrad
│   │   └── gradcam.py      # GradCAM
│   ├── dataset/            # Dataset utilities
│   │   ├── oxfordpet_dataset.py
│   │   ├── oxfordflower_dataset.py
│   │   └── miniimagenet_dataset.py
│   ├── classifier/         # Classifier utilities
│   ├── diffusion/          # Diffusion model components
│   ├── metric/             # Evaluation metrics
│   │   ├── diffid.py       # DiffID metric
│   │   ├── complexity.py   # Complexity metrics
│   │   └── path_stability.py
│   ├── mar_vae/            # MAR-VAE implementation
│   └── mig_vae/            # MIG-VAE implementation
├── configs/                # Configuration files
│   ├── benchmark/          # Benchmark configs for each method
│   │   ├── ig.yaml
│   │   ├── gig.yaml
│   │   ├── agi.yaml
│   │   ├── diffig.yaml
│   │   ├── mig.yaml
│   │   └── ...
│   ├── classifier/         # Classifier training configs
│   └── diffig/             # DiffIG specific configs
├── pipelines/              # Execution pipelines
│   ├── benchmark/          # Benchmark evaluation scripts
│   │   ├── diffid.py
│   │   ├── save_attributions.py
│   │   └── eval_attributions.py
│   ├── classifier/         # Classifier training pipelines
│   └── diffig/             # DiffIG training pipelines
├── notebooks/              # Jupyter notebooks
│   ├── cvpr26/             # CVPR 2026 figures
│   └── analysis notebooks
├── results/                # Experiment results
│   ├── benchmark/          # Benchmark results
│   ├── attributions/       # Saved attributions
│   ├── classifier_*/       # Trained classifiers
│   └── vae_*/              # Trained VAE models
├── scripts/                # Shell scripts
│   ├── benchmark_cvpr26.sh
│   ├── analyze_baseline_results.py
│   └── analyze_diffig_results.py
├── papers/                 # Related papers
└── requirements.txt        # Python dependencies

Key Components

Explainer Methods

• IG: Integrated Gradients
• GIG: Guided Integrated Gradients
• AGI: Adversarial Gradient Integration
• DiffIG: Diffusion-based IG (our method)
• MIG: Manifold Integrated Gradients
• BIG: Boundary Integrated Gradients
• EIG: Enhanced Integrated Gradients
• SG: SmoothGrad
• SPI: SmoothGrad Path Integration
• FullGrad: Full-Gradient Saliency
• GradCAM: Gradient-weighted Class Activation Mapping

Datasets Supported

• Oxford-IIIT Pet Dataset (37 categories)
• Oxford Flower Dataset (102 categories)
• Mini-ImageNet (100 classes)

Models Supported

• ResNet18
• ResNet34
• VGG16
• Inception

Evaluation Metrics

• DiffID: Diffusion-based attribution quality metric
• Insertion/Deletion: Standard faithfulness metrics
• Complexity: Attribution complexity measures
• Path Stability: Path consistency metrics

Key Features

• Comprehensive benchmark of attribution methods
• Novel DiffIG method using diffusion models for improved paths
• Support for multiple datasets and architectures
• Extensive evaluation metrics including novel DiffID metric
• Reproducible experiments with configuration files

Citation

This is a research project for developing and benchmarking improved explainability methods for image classifiers, with focus on path-based attribution methods and diffusion models.