Model compression method and device

Abstract

The embodiment of the invention discloses a model compression method and device. The method comprises the steps of obtaining a to-be-compressed model and compression preference configuration for the to-be-compressed model; determining a compression algorithm component and a corresponding algorithm hyper-parameter value according to the model type and the compression preference configuration of a model to be compressed, and using the compression algorithm component and the algorithm hyper-parameter value to carry out first compression on the to-be-compressed model to obtain a candidate compression result corresponding to the first compression; if the coincidence degree of the performance parameter of the candidate compression result corresponding to the first compression and the compressionpreference configuration does not meet the preset condition, executing the second compression; and continuing to generate a parameter adjustment strategy to adjust the compression algorithm componentand the algorithm hyper-parameter value used by the next compression until the coincidence degree of the performance parameter of the candidate compression result corresponding to the certain compression and the compression preference configuration meets a preset condition. A compression algorithm does not need to be adjusted manually, so that the influence caused by human experience is avoided,and the compression efficiency is improved.

Description

technical field [0001] The present application relates to the field of data processing, and in particular, to a model compression method and device. Background technique [0002] The network model can be deployed on different types of processing devices, and the processing device can implement specific functions, such as image recognition, data classification, etc., through the deployed network model. [0003] In some cases, however, the network model needs to be compressed prior to deployment in processing devices. Part of the reason is that the performance parameters of some network models are not good. For example, the model occupies a large space, the computing performance is not high, and the running speed is low. Such network models are generally network models designed based on human experience, or those with insufficient development experience. The network model developed by the developer. When this kind of network model is deployed to processing equipment, it not ...

AI Technical Summary

Problems solved by technology

Or, another part of the reason is that some processing devices that need to deploy network models do not have high processing capabilities, and it is difficult to support the deployment of network models with high-performance parameters. Therefore, it is necessary to compress such network models without losing too many performance parameters. Under the condition that it can be deployed in processing equipment with low proce...

Images