Grading deep learning model detection method and device for rare content, and computer equipment

Abstract

The invention discloses a grading deep learning model detection method and device for rare content and computer equipment. The method comprises the steps: adopting a hierarchical deep learning model mode, and configuring an inference model into a preliminary screening model and a decision-making model, wherein the preliminary screening model is a model meeting a preset performance threshold value,the decision-making model is a model meeting a preset index threshold value; adjusting the preliminary screening threshold value of the preliminary screening model to be a first threshold value, andaccording to the first threshold value, screening out normal content from the rare content, adjusting a decision threshold of the decision model to be a second threshold, and according to the second threshold, deciding out the normal content from the rare content after the normal content is screened out, wherein the second threshold value is less than the first threshold value. By means of the mode, sufficient reasoning performance indexes can be maintained, meanwhile, calculation resources needed for reasoning calculation can be reduced, and the machine cost is reduced.

Description

Technical field[0001]The present invention relates to the technical field of scarce content, in particular to a method, device and computer equipment for detecting a hierarchical deep learning model of scarce content.Background technique[0002]In the field of model checking, the current use of deep learning technology has been able to improve the recognition accuracy to a level beyond humans. Thanks to the rapid development of GPU (Graphics Processing Unit, graphics processing unit), through distributed large-scale training, the magnitude of model training time has been reduced from months or even years to days.[0003]Once model training is completed, inference deployment is required. Industrialized inference usually requires low latency and high concurrency. In order to maintain sufficient inference performance indicators, a large amount of computing resources are usually required for inference calculations, and machine costs account for the majority The cost of services.[0004]Differ...

AI Technical Summary

Problems solved by technology

[0005] However, the existing rare content detection schemes usually require a large amount of computing resources for inference calculations in order to maintain sufficient inference performance indicators, since the normal content accounts for the vast majority of the real data and the rare content accounts for a small proportion. cost of services

Images