US 12,443,661 B2
Model processing method and related device
Jingwei Huang, Wuhan (CN); Mingwei Sun, Wuhan (CN); and Shan Huang, Wuhan (CN)
Assigned to Huawei Technologies Co., Ltd., Shenzhen (CN)
Filed by HUAWEI TECHNOLOGIES CO., LTD., Guangdong (CN)
Filed on May 11, 2023, as Appl. No. 18/315,844.
Application 18/315,844 is a continuation of application No. PCT/CN2021/104960, filed on Jul. 7, 2021.
Claims priority of application No. 202011273450.X (CN), filed on Nov. 13, 2020.
Prior Publication US 2023/0281250 A1, Sep. 7, 2023
Int. Cl. G06F 16/90 (2019.01); G06F 16/901 (2019.01); G06F 30/20 (2020.01)
CPC G06F 16/9024 (2019.01) [G06F 30/20 (2020.01)] 20 Claims
OG exemplary drawing
 
1. A method, wherein the method comprises:
receiving, by a primary server in a server cluster, a first model sent by a client, wherein the server cluster includes the primary server and a plurality of subservers, the primary server includes at least one processor, each subserver includes an artificial intelligence (AI) processor, the first model is used to resolve a problem in at least one of three-dimensional modeling, image fusion, face reconstruction, or grid deformation, the first model comprises a plurality of first independent variables, and the plurality of first independent variables occupy a first quantity of computer memory resources that is greater than a threshold number;
generating, by the at least one processor of the primary server, a graph structure corresponding to the first model, wherein the first model comprises a plurality of first functions, the graph structure comprises a plurality of nodes and edges between the plurality of nodes, the nodes in the graph structure correspond to the first independent variables, and the edges between the plurality of nodes in the graph structure are determined based on the plurality of first functions;
randomly segmenting, by the at least one processor of the primary server, the graph structure to obtain at least two first sets, wherein each first set comprises at least one first independent variable;
sending, by the at least one processor of the primary server, the at least two first sets to at least two subservers of the plurality of subservers;
receiving, by the at least one processor of the primary server from the at least two subservers, a value of the at least one first independent variable in each first set of the at least two first sets;
updating, by the at least one processor of the primary server, the value of the at least one first independent variable in each first set by using a Levenberg-Marquardt (LM) algorithm, wherein updating the value of the at least one first independent variable in each first set comprises storing the at least one first independent variable in each first set using a second quantity of computer memory resources that is less than the threshold number;
repeatedly performing, by the at least one processor of the primary server, the random segmentation operation and the updating operation until a convergence condition of the first model is met to obtain solutions of the plurality of first independent variables in the first model; and
sending, by the at least one processor of the primary server, the solutions of the plurality of first independent variables in the first model to the client.