Hugging Face's logo

Datasets:
leekwoon
/
260204_sig_backup

Dataset card Files
xet
 Community
260204_sig_backup / README.md
leekwoon's picture
leekwoon
Upload README.md with huggingface_hub
aac938e verified
preview code
|
raw
history blame contribute delete
4.01 kB

Spectral IG Backup

Research project backup for Spectral Integrated Gradients (Spectral IG / SIG) - an explainability method that uses SVD-based coarse-to-fine path construction for image classifiers.

Download & Extract 

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

# 다운로드
huggingface-cli download leekwoon/260204_sig_backup --repo-type dataset --local-dir ./spectral_ig_data

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

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

Directory Structure 

spectral_ig/
├── cleanig/              # Core implementation modules
│   ├── explainer/        # Explainability algorithms
│   │   ├── ig.py         # Integrated Gradients
│   │   ├── spectral_ig.py # Spectral IG (main contribution)
│   │   ├── gig.py        # Guided IG
│   │   ├── big.py        # Blur IG
│   │   ├── agi.py        # Adversarial GI
│   │   ├── eig.py        # Enhanced IG
│   │   ├── mig.py        # Manifold IG
│   │   ├── samp.py       # SAMP
│   │   └── path_utils.py # Path generators (Linear, Spectral, Blur, Geodesic)
│   ├── dataset/          # Dataset utilities (ImageNet, OxfordPet, CIFAR10)
│   ├── classifier/       # Classifier utilities
│   └── metric/           # Evaluation metrics
├── configs/              # Hydra configuration files
│   ├── dataset/          # Dataset configs
│   ├── ig.yaml           # IG config
│   ├── spectral_ig.yaml  # Spectral IG config
│   ├── gig.yaml          # Guided IG config
│   └── ...
├── results/              # Experiment results
│   ├── icml26_figures/   # Paper figures
│   ├── roar/             # ROAR evaluation results
│   ├── diffid/           # DiffID evaluation results
│   ├── saliency_maps/    # Generated saliency maps
│   └── classifier_*/     # Trained classifiers
├── scripts/              # Execution scripts
│   ├── icml26_figures/   # Figure generation scripts
│   ├── roar.py           # ROAR evaluation
│   ├── diffid.py         # DiffID evaluation
│   └── measure_computation_time.py
├── saliency/             # Pre-computed saliency maps
└── requirements.txt      # Python dependencies

Key Components

Explainers

  • IG (Integrated Gradients): Standard linear path integration
  • Spectral IG: SVD-based coarse-to-fine path construction (main contribution)
  • GIG (Guided IG): Gradient-guided adaptive paths
  • BIG (Blur IG): Gaussian blur-based paths
  • AGI (Adversarial GI): Adversarial gradient integration
  • EIG (Enhanced IG): Latent space paths
  • MIG (Manifold IG): Geodesic paths on VAE manifold
  • SAMP: Adaptive sampling method

Spectral IG Key Parameters

  • num_steps: Number of integration steps (default: 200)
  • overlap: Overlap parameter for SVD component activation (default: 0.4)
    • overlap=1.0: All SVs scale together (like linear IG)
    • overlap=0.5: Partial overlap, skeleton first
    • overlap=0.0: Sequential, hard thresholding

Datasets

  • ImageNet (ILSVRC2012)
  • Oxford-IIIT Pet Dataset
  • CIFAR-10

Evaluation Metrics

  • ROAR (RemOve And Retrain): Measures attribution faithfulness
  • DiffID (Insertion/Deletion): Measures attribution quality
  • Computation time benchmarks

Usage 

# Install
pip install -e .

# Run Spectral IG
python scripts/diffid.py --config-name spectral_ig dataset=imagenet

# Run ROAR evaluation
python scripts/roar.py --config-name spectral_ig dataset=oxfordpet

# Generate paper figures
python scripts/icml26_figures/compare_ig_sig_frequency.py --config-name ig dataset=oxfordpet

Citation

Spectral Integrated Gradients: SVD-based Coarse-to-Fine Path Construction for Explainable AI (ICML 2026 submission)