Adaptive quantization method adaptive to neural network accelerator running on FPGA

Abstract

The invention discloses an adaptive quantization method adaptive to a neural network accelerator running on an FPGA, and belongs to the field of neural networks. The overflow degree of the neural network accelerator during calculation is automatically pre-judged according to the actual bit width in the calculation process of the neural network accelerator, the quantization parameters are adaptively adjusted according to the overflow degree, the problem of data overflow in the calculation process of a neural network algorithm on an FPGA is avoided, and therefore the correctness of a neural network model result is guaranteed. According to the adaptive quantification method, quantification operation is combined with resource planning of neural network accelerator hardware, the correctness of a result when the neural network accelerator deploys an algorithm is ensured, the model scale can be effectively reduced on the premise that the model precision and the execution efficiency are not lost, the method is easy to deploy and implement under the condition that resources are limited, storage space and computing resources are saved, and the method has important research significance and application value.

Description

technical field [0001] The invention belongs to the field of neural networks, in particular to an adaptive quantization method adapted to a neural network accelerator running on FPGA. Background technique [0002] In order to achieve high-speed and low-power calculations, neural network accelerators generally support low numerical precision, such as 8-bit or 6-bit fixed-point calculations, while the original numerical precision of neural network models is generally 32-bit floating-point numbers. Therefore, when a neural network accelerator deploys a neural network algorithm, it is necessary to automatically compress the neural network model into an 8-bit or 6-bit integer network through a quantization operation. [0003] The foreign giant Nvidia dominates the neural network accelerator market based on its complete GPU+CUDA ecology, and maps floating-point data to integer data. However, its products are expensive, do not have independent controllability, and GPU computing per...

AI Technical Summary

Problems solved by technology

[0004] The purpose of the present invention is to overcome the existing quantization method that does not consider the overflow problem that may occur during the integer data operation process when deployed on the FPGA, and cannot ensure the correctness of the calculation results, and provides a method that is adapted to run on FPGA Adaptive Quantization Methods for Neural Network Accelerators on

Images