System and method for executing deep neural network based on heterogeneous platform

Abstract

The invention provides a system and a method for executing a deep neural network based on a heterogeneous platform, which are used for solving the technical problems of low speed of executing the deepneural network based on the heterogeneous platform and poor compatibility with a model in the prior art, and are implemented by the following steps that: (1) a model analysis module analyzes a DNN model; (2) the analysis module constructs a calculation graph according to an analysis result; (3) a platform detection module detects hardware parameters of a heterogeneous computing unit in the heterogeneous platform; (4) a task allocation module constructs a task allocation strategy and optimizes the task allocation strategy; and (5) a reasoning module performs task allocation on the heterogeneous computing unit according to the optimal task allocation strategy to obtain a result of executing the DNN model. The analysis module supports analysis of the model format of the mainstream deep learning framework, the compatibility of the system to different framework models is improved, the task allocation module provides an optimal task allocation strategy, and the execution speed of the deep neural network is improved.

Description

technical field [0001] The invention belongs to the technical field of deep neural networks, and relates to a system and method for executing deep neural networks, in particular to a system and method for executing deep neural networks based on heterogeneous platforms, which can be used for target detection, face recognition and voice recognition, etc. field. Background technique [0002] With the rapid development of artificial intelligence, deep learning technology represented by deep neural network has achieved remarkable results. Deep neural network technology has been widely used in the fields of target detection, face recognition and speech recognition. With the rise of the field of big data, the demand for computing power of hardware devices by deep neural networks is also increasing. The training phase of the deep neural network model needs to be fed with a large number of training data sets. During the training cycle, it is usually necessary to perform hundreds of ...

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to overcome the deficiencies of the above-mentioned prior art, and propose a system and method for executing a deep neural network based on a heterogeneous platform, which is used to solve the problem of low execution speed of the neural network and compatibility with models existing in the prior art Poor technical issues

Images