Multi-precision neural network model implementation method and system

Abstract

The invention provides a multi-precision neural network model implementation method and system; the method comprises the steps: obtaining operator constraint conditions of a hardware accelerator, and generating calculation graph operator configuration information of a multi-precision neural network model; performing structural analysis on the multi-precision neural network model, and obtaining a corresponding relation between different numerical precision and a computational graph operator by combining configuration information of the computational graph operator; adjusting parameters of the multi-precision neural network model according to a corresponding relation between different numerical precision and computational graph operators, and obtaining performance compensation of the multi-precision neural network model after the numerical precision is reduced; and generating configuration information of each level according to the multi-precision neural network model after parameter adjustment, and deploying the multi-precision neural network model to a hardware accelerator according to the configuration information of each level. According to the method, when the complex neural network is deployed to a specific hardware accelerator, operators are ensured to adopt different numerical precision combinations, so that optimized power consumption, efficiency and storage bandwidth are obtained, and the model deployment difficulty is reduced.

Description

technical field [0001] The present invention relates to the field of computer technology, in particular to a method and system for realizing a multi-precision neural network model. Background technique [0002] Due to the high computational complexity of neural networks for complex tasks, multi-precision numerical representation methods are required on specific hardware accelerators to obtain optimized power consumption and efficiency. This requirement increases the difficulty of model deployment. [0003] In the prior art, the Chinese patent application with the publication number CN110942139A discloses a "Deep Learning Neural Network Deployment System and Its Method". Although this deployment method can support the identification, analysis, and deployment model generation of various neural network frameworks , but cannot support the implementation of multiple precision models on specific hardware accelerators. Contents of the invention [0004] In view of the defects in...

AI Technical Summary

Problems solved by technology

[0002] Due to the high computational complexity of neural networks for complex tasks, multi-precision numerical representation methods are required on specific hardware accelerators to obtain optimized power consumption and efficiency. This requirement increases the difficulty of model deployment

[0003] In the prior art, the Chinese patent application with the publication number CN110942139A discloses a "Deep Learning Neural Network Deployment System and Its Method". Although this deployment method can support the identification, analysis, and deployment model generation of various neural network frameworks , but cannot support the implementation of multiple precision models on specific hardware accelerators

Images