Model conversion method, apparatus and device, and computer storage medium

Abstract

The invention provides a model conversion method, apparatus and device, and a computer storage medium. The method comprises the steps of obtaining a to-be-converted model; determining neural network layers contained in the to-be-converted model, and obtaining conversion results corresponding to the neural network layers; and combining the conversion results corresponding to the neural network layers to obtain a conversion result corresponding to the to-be-converted model. According to the method, the model conversion cost can be reduced, and the model conversion efficiency is improved.

Description

【Technical field】 [0001] The present invention relates to the technical field of the Internet, in particular to a model conversion method, device, equipment and computer storage medium. 【Background technique】 [0002] In existing technologies, most machine learning models or deep learning models are trained and deployed on cloud servers. If you want to convert the learning model deployed in the cloud server into other types of models, such as converting them into models that can be applied to mobile devices, there are generally two ways: to convert the model as a whole through cross-compilation or based on training data Retrain the model. [0003] Among them, the method of converting the model through cross-compilation will have the following problems: developers need to have professional cross-compilation experience, and need to face various error reports and code modifications, so the conversion cost is high; if the model to be converted is not the same as The consistenc...

AI Technical Summary

Problems solved by technology

[0003] Among them, the method of converting the model through cross-compilation will have the following problems: developers need to have professional cross-compilation experience, and need to face various error reports and code modifications, so the conversion cost is high; if the model to be converted is not the same as The consistency between the models to be converted is not high, and the conversion will fail, so the conversion efficiency is low

However, the method of retraining the model based on the training data will have the following problems: the developer needs to discard the cloud server model that has been repeatedly tuned and business verified, and retrain the model on the new framework, which takes a lot of time and requires re-training. The performance of the trained model is not good enough compared to the performance of the model in the cloud server

Images