How evitanrelta/ml-bound-solver works
Main Components
a) Input Processing Layer:
• Input handlers for different model types (ONNX, PyTorch)
• Preprocessing pipeline from src/preprocessing/
• Conversion utilities (hwc_to_chw, linearize_conv2d)
b) Core Solver Component:
• Main solver architecture from src/modules/
• Solver layers hierarchy (input, intermediate, output layers)
• AdversarialCheckModel component
c) Training System:
• Training infrastructure from src/training/
• EarlyStopHandler and TrainingConfig components
d) Comparison System:
• Gurobi comparison functionality
• Result validation components
Level 1: Input Layer (Model & Configuration Intake)
Level 2: Preprocessing Layer
Level 3: Core Solver Components
Level 4: Training & Optimization Layer
Level 5: Output & Validation Layer
Component Relationships
• Model input → Preprocessing → Solver
• Solver → Training system
• Results → Gurobi comparison
• Configuration files → Various components
Specific Implementation Details
Input Handlers:
• ONNX Model Loader
• PyTorch Model Loader
• Configuration Parser
Preprocessing:
• Linearization Module
• Conversion Utilities
• Solver Input Builder
Core Solver:
• SolverSequential
• Layer Handlers (Input/Intermediate/Output)
• AdversarialCheckModel
Training:
• Training Configuration
• Early Stop Handler
• Training Loop
Validation:
• Gurobi Comparator
• Results Analyzer
Technical Details
Annotations for:
• GPU/CPU computation support
• PyTorch integration points
• ONNX model handling
• Gurobi interface points
Additional Elements
Include:
• Configuration flow (YAML files)
• Device management (CPU/GPU)
• Data structure transformations
• Optimization paths