How b-cos/b-cos-v2 works
Main Components
a) Core Architecture Components:
• Model Architecture Layer
• Base Models (ResNet, DenseNet, VGG, ConvNext, ViT variants)
• B-cos Specific Modules (BCOSConv2d, BCOSLinear)
• Normalization Modules (Centered and Uncentered)
• Data Processing Layer
• Data Loading & Caching System
• Transforms Pipeline
• Sampling Mechanisms
• Training Infrastructure
• Training Loop Management
• Optimization Components
• Callbacks System
• Interpretability Framework
• Explanation Methods
• Analysis Tools
• Visualization Components
b) Supporting Components:
• Configuration Management
• Experiment Management
• Model Zoo Interface
• Evaluation Pipeline
Component Relationships and Data Flow
[Input Data] → [Data Processing Layer]
• ↓
[Model Architecture Layer] ← [B-cos Specific Modules]
• ↓
[Training Infrastructure] ← [Optimization Components]
• ↓
[Interpretability Framework]
• ↓
[Evaluation & Analysis]
Specific Implementation Details
• Rectangles: Core processing modules
• Hexagons: Data storage/caching components
• Diamonds: Decision points/configuration
• Rounded rectangles: External interfaces
• Blue: Core model components
• Green: Data processing
• Orange: Training infrastructure
• Purple: Interpretability components
• Gray: Utility/support components
Key Architectural Patterns to Highlight
a) Modular Design:
• Clear separation between model definitions, training infrastructure, and interpretability components
• Plugin architecture for different model types
b) Data Flow:
• Caching mechanism with Redis/SHM
• Transform pipeline flow
c) Training Pipeline:
• Demonstrate the interaction between trainer, callbacks, and metrics tracking
• Optimization loop with learning rate scheduling
• Data Pipeline |
• CachedImageFolder |
• Data Transforms |
• Sampling Logic |
• Model Components |
• Base Networks |
• B-cos Modules |
• Normalization |
• Training Pipeline |
• Trainer |
• Callbacks |
• Optimization |
• Interpretability |
• Explainers |
• Analysis Tools |
• Visualization |