Distributed model training optimization method and device, equipment, medium and program product

Abstract

The invention provides a distributed model training optimization method and device, network equipment, a medium and a program product, and relates to the technical field of model training. The distributed model training optimization method comprises the following steps: sending model parameters required for training a risk prevention and control model to a specified working node in a parameter server system; receiving a first mapping table sent by the specified working node, the first mapping table being a mapping table between a parameter identifier of the model parameter and a compression gradient, and the compression gradient being obtained by compressing a risk feature gradient of the model parameter by the specified working node; counting the model parameters based on the first mapping table; and when detecting that the access frequency of the model parameter is greater than a frequency threshold based on a counting result, processing the model parameter based on a load balancing rule. Through the technical scheme disclosed by the invention, the network bandwidth bottleneck can be relieved.

Images