🤖 Model Interpretability Framework

A deep learning model visualization and interpretation framework that helps understand model behavior through interactive visualizations.

🌟 Overview

This framework provides tools to analyze and visualize:

• 🧮 Model attention patterns
• 📊 Token importance scores
• 🎯 Uncertainty landscapes
• 🔮 Embedding spaces
• ⚡ Layer dynamics
• 🎲 Decision boundaries

✨ Key Features

📊 1. Basic Visualizations

• 📈 Token importance bar charts
• 🔥 Attention heatmaps
• 📉 Confidence scores

🔬 2. Advanced Analysis

• 🌌 2D embedding space visualization
• 🔄 Cross-model comparisons
• 📱 Interactive plots with Plotly

🚀 3. Deep Model Analysis

• 🔄 Layer-wise representation dynamics
• 📊 Multi-metric visualization
• ⚡ Attention flow patterns

🎯 4. Uncertainty Analysis

• 🌐 3D uncertainty landscapes
• 📈 Predictive entropy visualization
• 📉 Confidence evolution plots

🚀 Quick Start

1. Install the package:

git clone https://github.com/Parisaroozgarian/model-interpretability.git
cd model-interpretability
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate
pip install -r requirements.txt
pip install -e .

2. Basic usage:

from src.config.visualization_config import VisualizationConfig
from src.visualizers.feature_importance import FeatureImportanceVisualizer

# Initialize
config = VisualizationConfig()
viz = FeatureImportanceVisualizer(config)

# Create visualization
fig = viz.plot_token_importance(
    tokens=["This", "is", "a", "test"],
    importance_scores=[0.3, 0.2, 0.1, 0.4]
)
fig.show()

📚 Examples

Check the examples/ directory for more usage examples:

• 📊 basic_visualization.py: Simple token importance and uncertainty plots
• 🔬 advanced_visualization.py: 2D embedding space and attention patterns
• 🚀 deep_analysis.py: Layer dynamics and model behavior analysis

📁 Project Structure

model-interpretability/
├── src/                 # Source code
│   ├── config/         # ⚙️ Configuration settings
│   │   └── visualization_config.py
│   ├── visualizers/    # 📊 Visualization modules
│   │   ├── base.py
│   │   ├── feature_importance.py
│   │   ├── uncertainty.py
│   │   ├── advanced.py
│   │   └── deep_analysis.py
│   └── utils/          # 🛠️ Utility functions
├── examples/           # 📚 Usage examples
├── tests/             # 🧪 Test suite
└── docs/              # 📖 Documentation

🛠️ Dependencies

torch>=1.9.0
transformers>=4.5.0
plotly>=5.1.0
numpy>=1.19.5
seaborn>=0.11.2
tensorflow-hub>=0.12.0
nltk>=3.6.3
umap-learn>=0.5.1
scikit-learn>=0.24.2
networkx>=2.6.2

👨‍💻 Development

Run tests:

python -m pytest tests/

Code coverage:

pytest tests/ --cov=src/ --cov-report=xml

👩‍💻 Author

Parisa Roozgarian

🙏 Acknowledgments

• 🤗 BERT model from Hugging Face Transformers
• 📊 Plotly for interactive visualizations
• 🌐 UMAP for dimensionality reduction

📝 License

This project is licensed under the MIT License - see the LICENSE file for details