DeepBridge Distillation

Model Compression and Knowledge Distillation Toolkit - Extension for DeepBridge

Part of the DeepBridge v2.0 Ecosystem

This package was extracted from DeepBridge v1.x to provide focused model compression capabilities. See Migration Guide if migrating from v1.x.

Installation

pip install deepbridge-distillation

This will automatically install deepbridge>=2.0.0 as a dependency.

Quick Start

from deepbridge import DBDataset
from deepbridge_distillation import AutoDistiller

# Create dataset with teacher model predictions
dataset = DBDataset(
    data=df,
    target_column='target',
    features=features,
    prob_cols=['prob_0', 'prob_1']
)

# Run automated distillation
distiller = AutoDistiller(
    dataset=dataset,
    output_dir='results',
    n_trials=10
)
results = distiller.run(use_probabilities=True)

Features

• Automated Distillation: AutoDistiller with hyperparameter optimization
• Knowledge Distillation: Transfer knowledge from teacher to student models
• Surrogate Models: Create efficient surrogate models
• HPM Knowledge Distillation: Hierarchical Prototype-based Method
• Multi-framework Support: Works with scikit-learn, XGBoost, PyTorch

Documentation

Full documentation: https://deepbridge.readthedocs.io/en/latest/distillation/

Related Projects

• deepbridge - Model Validation Toolkit (core)
• deepbridge-synthetic - Synthetic Data Generation

License

MIT License - see LICENSE