Fixed-point quantification method and device for deep learning model

Abstract

The invention discloses a fixed-point quantification method and device for a deep learning model, and the method comprises the following steps: inputting calibration data to a target model, taking a model parameter and an activation value of the target model as quantification objects in sequence, and executing the following steps: inputting calibration set data, extracting the quantification objects of the target model according to layers, acquiring a distribution histogram of a quantized object, scaling the distribution histogram of the quantized object through an adaptive KL divergence equation, acquiring KL divergence values corresponding to different decimal point positions based on a preset quantization digit, and comparing to obtain a first quantization result of the quantized object. The defect that the KL divergence algorithm only pays attention to the probability is overcome, the quantization result is optimized, and the quantization speed can be greatly increased, the quantization efficiency is improved and the time is saved under the condition that the certain precision of the quantized model is ensured.

Description

technical field [0001] The invention relates to the technical field of deep learning, in particular to a method and device for fixed-point quantization of a deep learning model. Background technique [0002] Common deep learning models usually have a large number of parameters and layers in order to improve network accuracy, resulting in a sharp increase in model storage size and slow inference speed. Slow reasoning makes many high-precision deep learning networks can only run on GPU systems with high computing power, making it difficult to implement applications. [0003] With the gradual warming up of deep learning in recent years, its deployment requirements are increasing, and the end-side often uses low-precision computing units in order to balance cost and speed. Compared with full-precision or mixed-precision models including half-precision, although low-precision, it can reason faster and occupy less memory. Therefore, how to convert between full-precision and low-p...

AI Technical Summary

Problems solved by technology

[0002] Common deep learning models usually have a large number of parameters and layers in order to improve network accuracy, resulting in a sharp increase in model storage size and slow inference speed

Slow reasoning makes many high-precision deep learning networks can only run on GPU systems with high computing power, making it difficult to implement applications

Images