Representative Input Set Selection for ANN Quantization Calibration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining a representative input data set for post-training quantization of artificial neural networks are inefficient, time-consuming, and often result in suboptimal accuracy due to the need for subjective manual selection and numerous trial-and-error cycles, especially when deploying on resource-constrained embedded systems.
Innovation Solution
A computer-implemented method that automatically determines a representative input data set by extracting parameters, setting clipping limits, and iteratively adapting them based on quantization error to identify a subset of the input data set that best represents the network inputs, reducing the size of the data set required for quantization calibration.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual selection and trial-and-error cycles are used to determine representative input data set, then accuracy can be improved, but time consumption increases significantly
Solution Approach 1:
The system automatically determines the representative input data set by extracting parameters and adaptively selecting data points based on quantization error metrics, eliminating the need for manual selection and trial-and-error cycles while maintaining high accuracy
Solution Approach 2:
The method uses feedback from quantization error calculations to iteratively refine the selection of representative data points, allowing the system to self-optimize the calibration process without human intervention and significantly reduce time consumption
2Measurement precision
If large input data set is used for quantization calibration, then accuracy is maintained, but computational resources and time increase
Solution Approach 1:
The system extracts and selects only the most representative data points from the large input data set by evaluating quantization error metrics, creating a reduced calibration set that maintains accuracy while significantly improving computational efficiency
Solution Approach 2:
The method dynamically adjusts parameters such as clipping limits and data selection criteria based on quantization error calculations, enabling the system to identify optimal representative data points that balance accuracy and efficiency
3Ease of operation
If existing quantization methods are applied without model access, then ease of operation improves, but accuracy deteriorates
Solution Approach 1:
The system introduces an intermediary parameter extraction and evaluation mechanism that operates on input data sets without requiring direct access to the ANN model, using quantization error metrics as a mediator to determine representative data points while maintaining both ease of operation and accuracy
Data Source
AI summary
A computer implemented method for determining a representative input data set for post-training quantization of artificial neural networks comprises: acquiring an input data set; processing the input data set to extract one or more parameters associated with the input data set; setting clipping limits comprising a lower clipping limit and an upper clipping limit; adapting the upper and lower clipping limits by computing (656) a quantization error based on a numerical difference between the input data set and a quantized and a dequantized input data set, wherein the error is computed based on an error metric which is based on user input; and determining a representative input data set as a subset of the input data set whose elements fit within the clipping limits; wherein the representative input data set is to be used for quantization calibration of the post-training quantization.