Domain sensitive unloading method for deep reasoning service in edge network

The invention relates to a domain sensitive unloading method oriented to a deep reasoning service in an edge network. In order to solve the problem that a deep learning model of domain specific knowledge deployed on an edge server has significant difference on performance expressions of different reasoning tasks, the invention provides a method for selecting and deploying the edge server of a model domain most suitable for the reasoning tasks by fully utilizing diversity of the edge server model domain. Therefore, the unloading performance is improved. The method comprises the following steps: designing a field-sensitive unloading-oriented performance index for describing the sensitivity degrees of different learning model fields and unloading tasks to the model fields; designing an efficient retrieval mechanism, and retrieving information of a learning model deployment field on the edge server so as to determine the edge server providing inference service; and a calculation unloading algorithm oriented to th