leekwoon commited on
Commit
47ea0b6
·
verified ·
1 Parent(s): 90163f7

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +127 -0
README.md ADDED
@@ -0,0 +1,127 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # DiffIG GIG Backup
2
+
3
+ Research project backup for DiffIG (Diffusion-based Integrated Gradients) and GIG (Guided Integrated Gradients) - advanced explainability methods for image classifiers.
4
+
5
+ ## Download & Extract
6
+
7
+ ```bash
8
+ # huggingface_hub 설치 (필요시)
9
+ pip install huggingface_hub
10
+
11
+ # 다운로드
12
+ huggingface-cli download leekwoon/diffig_gig_backup --repo-type dataset --local-dir ./diffig_data
13
+
14
+ # 무결성 확인 (선택사항)
15
+ cd diffig_data
16
+ md5sum -c checksums.md5
17
+
18
+ # 파일 합치기 및 압축 해제
19
+ cat data.tar.gz.part_* | tar -xzvf -
20
+ ```
21
+
22
+ ## Directory Structure
23
+
24
+ ```
25
+ diffig/
26
+ ├── diffig/ # Core implementation modules
27
+ │ ├── explainer/ # Explainability algorithms
28
+ │ │ ├── ig.py # Integrated Gradients
29
+ │ │ ├── ig2.py # Improved IG
30
+ │ │ ├── gig.py # Guided IG
31
+ │ │ ├── agi.py # Adversarial GI
32
+ │ │ ├── diffig.py # DiffIG (our method)
33
+ │ │ ├── mig.py # Manifold IG
34
+ │ │ ├── big.py # Boundary IG
35
+ │ │ ├── eig.py # Enhanced IG
36
+ │ │ ├── sg.py # Smooth Gradients
37
+ │ │ ├── spi.py # SmoothGrad Path Integration
38
+ │ │ ├── fullgrad.py # FullGrad
39
+ │ │ └── gradcam.py # GradCAM
40
+ │ ├── dataset/ # Dataset utilities
41
+ │ │ ├── oxfordpet_dataset.py
42
+ │ │ ├── oxfordflower_dataset.py
43
+ │ │ └── miniimagenet_dataset.py
44
+ │ ├── classifier/ # Classifier utilities
45
+ │ ├── diffusion/ # Diffusion model components
46
+ │ ├── metric/ # Evaluation metrics
47
+ │ │ ├── diffid.py # DiffID metric
48
+ │ │ ├── complexity.py # Complexity metrics
49
+ │ │ └── path_stability.py
50
+ │ ├── mar_vae/ # MAR-VAE implementation
51
+ │ └── mig_vae/ # MIG-VAE implementation
52
+ ├── configs/ # Configuration files
53
+ │ ├── benchmark/ # Benchmark configs for each method
54
+ │ │ ├── ig.yaml
55
+ │ │ ├── gig.yaml
56
+ │ │ ├── agi.yaml
57
+ │ │ ├── diffig.yaml
58
+ │ │ ├── mig.yaml
59
+ │ │ └── ...
60
+ │ ├── classifier/ # Classifier training configs
61
+ │ └── diffig/ # DiffIG specific configs
62
+ ├── pipelines/ # Execution pipelines
63
+ │ ├── benchmark/ # Benchmark evaluation scripts
64
+ │ │ ├── diffid.py
65
+ │ │ ├── save_attributions.py
66
+ │ │ └── eval_attributions.py
67
+ │ ├── classifier/ # Classifier training pipelines
68
+ │ └── diffig/ # DiffIG training pipelines
69
+ ├── notebooks/ # Jupyter notebooks
70
+ │ ├── cvpr26/ # CVPR 2026 figures
71
+ │ └── analysis notebooks
72
+ ├── results/ # Experiment results
73
+ │ ├── benchmark/ # Benchmark results
74
+ │ ├── attributions/ # Saved attributions
75
+ │ ├── classifier_*/ # Trained classifiers
76
+ │ └── vae_*/ # Trained VAE models
77
+ ├── scripts/ # Shell scripts
78
+ │ ├── benchmark_cvpr26.sh
79
+ │ ├── analyze_baseline_results.py
80
+ │ └── analyze_diffig_results.py
81
+ ├── papers/ # Related papers
82
+ └── requirements.txt # Python dependencies
83
+ ```
84
+
85
+ ## Key Components
86
+
87
+ ### Explainer Methods
88
+ - **IG**: Integrated Gradients
89
+ - **GIG**: Guided Integrated Gradients
90
+ - **AGI**: Adversarial Gradient Integration
91
+ - **DiffIG**: Diffusion-based IG (our method)
92
+ - **MIG**: Manifold Integrated Gradients
93
+ - **BIG**: Boundary Integrated Gradients
94
+ - **EIG**: Enhanced Integrated Gradients
95
+ - **SG**: SmoothGrad
96
+ - **SPI**: SmoothGrad Path Integration
97
+ - **FullGrad**: Full-Gradient Saliency
98
+ - **GradCAM**: Gradient-weighted Class Activation Mapping
99
+
100
+ ### Datasets Supported
101
+ - Oxford-IIIT Pet Dataset (37 categories)
102
+ - Oxford Flower Dataset (102 categories)
103
+ - Mini-ImageNet (100 classes)
104
+
105
+ ### Models Supported
106
+ - ResNet18
107
+ - ResNet34
108
+ - VGG16
109
+ - Inception
110
+
111
+ ### Evaluation Metrics
112
+ - **DiffID**: Diffusion-based attribution quality metric
113
+ - **Insertion/Deletion**: Standard faithfulness metrics
114
+ - **Complexity**: Attribution complexity measures
115
+ - **Path Stability**: Path consistency metrics
116
+
117
+ ## Key Features
118
+
119
+ - Comprehensive benchmark of attribution methods
120
+ - Novel DiffIG method using diffusion models for improved paths
121
+ - Support for multiple datasets and architectures
122
+ - Extensive evaluation metrics including novel DiffID metric
123
+ - Reproducible experiments with configuration files
124
+
125
+ ## Citation
126
+
127
+ 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.