leekwoon/latent_gig_backup

Latent GIG Backup

Research project backup for Latent Gradient Integrated Gradients (Latent GIG) - an explainability method for image classifiers that leverages latent space paths.

Download & Extract

# huggingface_hub 설치 (필요시)
pip install huggingface-hub==0.20.2

# 다운로드
huggingface-cli download leekwoon/latent_gig_backup --repo-type dataset --local-dir ./latent_gig_data

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

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

Directory Structure

latent_gig/
├── cleanig/              # Core implementation modules
│   ├── explainer/        # Explainability algorithms
│   │   ├── ig.py         # Integrated Gradients
│   │   ├── gig.py        # Guided IG
│   │   ├── agi.py        # Adversarial GI
│   │   ├── eig.py        # Expected IG
│   │   ├── mig.py        # Manifold IG
│   │   ├── saliency.py   # Saliency maps
│   │   ├── latent_gig.py # Latent GIG (main method)
│   │   ├── soft_latent_gig.py
│   │   ├── superpixel_gig.py
│   │   └── smooth_wrapper.py
│   ├── dataset/          # Dataset utilities (ImageNet, OxfordPet, OxfordFlower)
│   ├── classifier/       # Classifier utilities
│   ├── metric/           # Evaluation metrics (DiffID, DiffROAR)
│   ├── mar_vae/          # MAR-VAE implementation
│   └── vae_wrapper.py    # VAE wrapper for different models
├── configs/              # Configuration files
│   ├── dataset/          # Dataset configs
│   ├── ig.yaml           # Integrated Gradients config
│   ├── gig.yaml          # Guided IG config
│   ├── agi.yaml          # Adversarial GI config
│   ├── eig.yaml          # Expected IG config
│   ├── mig.yaml          # Manifold IG config
│   ├── latent_gig.yaml   # Latent GIG config
│   ├── smooth_*.yaml     # Smooth variants (grad, ig, gig, latent_gig)
│   ├── saliency.yaml     # Saliency config
│   └── superpixel_gig.yaml
├── data/                 # Dataset files
│   ├── cifar-10-*/       # CIFAR-10 dataset
│   └── cifar-100-*/      # CIFAR-100 dataset
├── results/              # Experiment results
│   ├── benchmark_diffid/ # DiffID benchmark results
│   ├── benchmark_diffroar/ # DiffROAR benchmark results
│   ├── classifier_oxfordpet/   # Trained OxfordPet classifiers
│   ├── classifier_oxfordflower/ # Trained OxfordFlower classifiers
│   ├── tests/            # Test results with path visualizations
│   └── icml26_*/         # ICML26 experiment results
├── scripts/              # Execution scripts
│   ├── benchmark_diffid/ # DiffID benchmark scripts
│   ├── benchmark_diffroar/ # DiffROAR benchmark scripts
│   ├── tests/            # Test scripts
│   ├── diffid.py         # DiffID evaluation
│   ├── diff_roar.py      # DiffROAR evaluation
│   ├── analyze_results.py # Result analysis tool
│   └── upload_to_hf.sh   # HuggingFace upload script
├── requirements.txt      # Python dependencies
└── setup.py              # Package setup

Key Components

Explainers

• IG (Integrated Gradients): Standard integrated gradients
• GIG (Guided IG): Guided version with adversarial paths
• AGI (Adversarial GI): Adversarial gradient integration
• EIG (Expected IG): Expected integrated gradients
• MIG (Manifold IG): Manifold-based integrated gradients
• Saliency: Basic saliency maps
• Latent GIG: Our method using latent space paths with VAE models
• Smooth variants: SmoothGrad, SmoothIG, SmoothGIG, SmoothLatentGIG

VAE Models Supported

• SD1: Stable Diffusion v1 VAE
• SD2: Stable Diffusion v2 VAE
• KD: Kandinsky VAE
• MAR: Masked Auto-Regressive VAE

Datasets

• ImageNet (subset)
• Oxford-IIIT Pet Dataset
• Oxford Flower Dataset
• CIFAR-10, CIFAR-100

Evaluation Metrics

• DiffID: Diffusion-based deletion metric
• DiffROAR: Diffusion-based ROAR metric
• Insertion/Deletion curves
• Attribution quality metrics

Parameters for Latent GIG

• fraction: Amount of latent to move at each step (0.05, 0.1, 0.2, 0.3)
• vae_type: VAE model to use (mar, sd1, sd2, kd)

Citation

This is a research project for developing improved explainability methods for image classifiers using latent space interpolation.