How linear95/club works
CLUB is a research-code repository for estimating mutual information and minimizing it in learning models. The main workflows are running the estimator comparisons, training an Information Bottleneck model, and training/evaluating a domain-adaptation model. The README documents the notebook capabilities; source excerpts show selected implementation details, while some links rely on the SOURCE INDEX or README rather than inspected notebook contents.
Main entry points: main_DANN.py DA Training and Evaluation · CLUB_MNIST_IB.py IB Training · utils.py DA Utilities · mi_estimation.ipynb Estimator Demo
Components 11
MI Estimation
- Estimator Demo
mi_estimation.ipynb - Gaussian Demo
mi_minimization.ipynb
Information Bottleneck
- IB Training
MI_IB/CLUB_MNIST_IB.py - IB Networks
MI_IB/architecture.py - IB Utilities
MI_IB/utils.py
Domain Adaptation
- Domain Data Loader
MI_DA/imageloader.py - DA Training and Evaluation
MI_DA/main_DANN.py - DANN Model
MI_DA/MNISTModel_DANN.py - DA Utilities
MI_DA/utils.py
Shared MI Methods
- MI Estimators
mi_estimators.py
Other components
- Researcher
Connections 12
- Researcher → Estimator Demo: runs comparison
README.md - Estimator Demo → MI Estimators: uses estimators
README.md - Researcher → Gaussian Demo: runs toy example
README.md - Researcher → IB Training: runs IB training
- IB Training → IB Networks: calls networks
MI_IB/CLUB_MNIST_IB.py - IB Training → IB Utilities: uses utilities
MI_IB/CLUB_MNIST_IB.py - Researcher → DA Training and Evaluation: runs adaptation
- DA Training and Evaluation → Domain Data Loader: loads datasets
MI_DA/main_DANN.py - DA Training and Evaluation → DANN Model: trains model
MI_DA/main_DANN.py - DA Training and Evaluation → DA Utilities: uses utilities
MI_DA/main_DANN.py - Domain Data Loader → DA Utilities: uses encoding
MI_DA/imageloader.py - DANN Model → DA Utilities: uses helpers
MI_DA/MNISTModel_DANN.py