How naforoutan/relafis works
ReLaFIS is a differentiable neuro-fuzzy classification framework. The principal workflow takes feature vectors through learned relational fuzzy terms, forms and normalizes rule activations, then produces class logits and a reconstruction used by an auxiliary loss. The repository also contains related neuro-fuzzy model implementations, dataset adapters, and a shared linguistic-richness/ARE utility. The README documents capabilities beyond the excerpts; those are represented where their implementation path is supported. Unsampled dataset-adapter wiring is not inferred.
Main entry points: ReLaFIS.py GIFT relations · linguistic_richness.py ARE and richness · litanfis.py LitANFIS model · griffin.py GRIFFIN model
Components 12
ReLaFIS inference
- GIFT relations
model/ReLaFIS.py - Rule activations
model/ReLaFIS.py - Input reconstruction
model/ReLaFIS.py - Local consequents
model/ReLaFIS.py - Class logits
model/ReLaFIS.py
Related models
- ANFIS baseline
model/anfis.py - UNFIS model
model/unfis.py - LitANFIS model
model/litanfis.py - GRIFFIN model
model/griffin.py
Linguistic analysis
- ARE and richness
utils/linguistic_richness.py
Experiment support
- Dataset adapters
tests
Other components
- Feature input
Connections 8
- Feature input → GIFT relations: provides features
README.md - GIFT relations → Rule activations: forms rules
model/ReLaFIS.py - Rule activations → Input reconstruction: reconstructs input
model/ReLaFIS.py - Rule activations → Local consequents: weights consequents
model/ReLaFIS.py - Local consequents → Class logits: produces logits
model/ReLaFIS.py - GIFT relations → ARE and richness: measures relations
model/ReLaFIS.py - LitANFIS model → ARE and richness: uses categories
model/litanfis.py - GRIFFIN model → ARE and richness: uses categories
model/griffin.py