How calad0i/quantizers works
Main Components
a) Core Quantization Modules:
• Fixed-Point Quantizer
• Minifloat Quantizer
• Binary Quantizer
• Stateful Quantizers
b) Implementation Variants:
• Keras Implementation
• NumPy Implementation
• C++ Reference Implementation
Top Level (API Layer):
• Stateless Quantizer Interface
• Stateful Quantizer Interface
Middle Level (Implementation Layer):
• Keras Backend
• NumPy Backend
• C++ Reference Implementation
Bottom Level (Core Operations):
• Binary Operations
• Fixed-Point Operations
• Minifloat Operations
Component Relationships
• How the API layer interfaces with different backend implementations
• Relationship between stateful and stateless quantizers
• Demonstrate the flow from high-level quantizer calls to low-level numerical operations
Specific Elements
a) Quantizer Components:
• Rectangle: Fixed-Point Quantizer
• Hexagon: Minifloat Quantizer
• Circle: Binary/Ternary Quantizer
• Diamond: Stateful Quantizers
b) Implementation Variants:
• Blue: Keras Implementation
• Green: NumPy Implementation
• Orange: C++ Reference
c) Data Flow:
Technical Details to Emphasize
• Relationship between Python implementations and C++ reference code
• Where gradients are computed (Keras implementation)
• Components that support stochastic rounding
• Bit-accurate precision matching process
Additional Notes
• Version compatibility information (Python ≥3.10, Keras ≥3.0)
• Hardware synthesis capabilities through HLS