Neural network transfer learning automatic training scheduling method based on ONNX model

Abstract

The invention relates to a neural network transfer learning automatic training scheduling method based on an ONNX model, and the method comprises the following steps: S1, building a PI database whichis used for the collection and storage of process data of a production environment; s2, respectively adopting different deep learning model training platforms to train the process data according to different task requirements, respectively exporting the obtained models as ONNX models, and forming an ONNX model library; s3, performing matching from the ONNX model library according to the task requirement of the production environment to obtain a corresponding target model; s4, deploying the target model to a corresponding edge device end; s5, enabling the edge equipment end to use the deployedtarget model to complete a corresponding task, and monitoring whether the performance of the target model is reduced or not; and S6, if so, loading new process data, and going to the step S2. According to the method, an automatic training scheduling method is utilized, the adaptability of the model to the production environment is improved, and the generalization performance of the model is better.

Description

technical field [0001] The invention belongs to the technical field of deep learning, and in particular relates to an automatic training scheduling method for neural network migration learning based on an ONNX model. Background technique [0002] At present, all major deep learning frameworks have their own unique characteristics and advantages, and each has its own set of proprietary model representation methods, and its own set of standards to operate and run its own model. Multiple learning frameworks lead to time-consuming and labor-intensive migration of algorithms and models between different frameworks. [0003] In addition, through the deep learning framework, the model derived in the training environment often does not work well in the production environment. Due to the conditions of the production environment or other factors, the model obtained in the training environment cannot fit the distribution of the data well, resulting in a decline in the performance of t...

AI Technical Summary

Problems solved by technology

Multiple learning frameworks lead to time-consuming and laborious migration of algorithms and models between different frameworks

[0003] In addition, through the deep learning framework, the model obtained in the training environment often does not work well in the production environment

Due to the conditions of the production environment or other factors, the model obtained in the training environment cannot fit the distribution of the data well, resulting in a decline in the performance of the model

Especially for data that did not appear during training, it may lead to changes in data distribution

Model generalization performance needs to be improved

Images