Module calibration 

Calibration utilities for quantization

This module provides utilities for finding optimal quantization parameters. Quantization calibration is the process of determining the optimal scaling factors and zero points for a given dataset and quantization method.

The module includes:

• Histogram-based methods for range calibration
• Entropy-based methods using KL divergence minimization
• Per-channel calibration strategies
• Dynamic calibration based on data statistics

Structs

CalibrationConfig

Configuration for quantization calibration

Enums

CalibrationMethod

Calibration method for determining quantization parameters

Functions

calibrate_matrix

Calibrate quantization parameters for a matrix using the specified method

calibrate_vector

Calibrate quantization parameters for a vector using the specified method

create_params_from_range

Create QuantizationParams from a min-max range

determine_data_type

Determine the appropriate data type based on bit width

find_min_max

Find the minimum and maximum values in a matrix

find_min_max_vec

Find the minimum and maximum values in a vector

get_activation_calibration_config

Helper function to get recommended calibration configuration for neural network activations

get_weight_calibration_config

Helper function to get recommended calibration configuration for neural network weights