AI model publishing method and system based on Seldon

Abstract

The invention discloses an AI model publishing method and system based on Seldon, and the method comprises the steps: constructing a docker mirror image based on Seldon, storing the docker mirror image in a mirror image warehouse of a mirror image management module, and enabling the docker mirror image to serve as a basic container for model publishing; and enabling the container management module to start a model container through the basic mirror image, input corresponding model parameters through a command line, automatically generating a model loading file model.py, downloading and installing an environment installation package required by a corresponding model from a product library, and downloading a corresponding model file from a model library. And when the model.py is operated, loading and releasing the model of the corresponding version. And finally, registering the published model to a gateway through a service management module, completing management of the published model by controlling registration information of the published model, and solving the problem of 'last kilometer' of continuous publishing of the model.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence, in particular to a Seldon-based AI model publishing method and system. Background technique [0002] At present, a common method of model release is to use TensorFlow Serving, but it only supports the release of models generated by the TensorFlow framework. Another commonly used model release method is to use the Seldon platform to release machine learning models. Seldon is an open-source platform for deploying machine learning models on Kubernetes. Seldon allows data scientists to create models using any machine learning toolkit or programming language, including mainstream TensorFlow models, Pytorch and other models. The Seldon platform exposes machine learning models via REST and gRPC when deploying models so that they can be integrated into desired business applications. And Seldon allows complex runtime reasoning graphs to be deployed as microservices, including model serv...

AI Technical Summary

Problems solved by technology

However, the loading files of common machine learning models have high similarity in content, and the model needs to be rewritten every time the model is loaded, which seriously affects the development efficiency of developers.

Moreover, for large-scale model deployment, it is necessary to write a model loading file every time it is used, and it is impossible to achieve continuous release of the model

Images