Multi-class model training method based on gradient balance, medium and equipment

Abstract

The invention provides a multi-class model training method based on gradient balance, a medium and equipment. The model training method comprises the following steps: inputting training sample data into a selected neural network model for training to obtain a loss function; performing statistics on gradient distribution of a loss function during model training; distributing sample weights according to a distribution result; performing weight smoothing processing; performing weight attenuation processing; obtaining a new gradient updating network model; according to the method, the weight of sample distribution is adjusted according to the distribution of the loss function gradient obtained by inputting the training sample data in the training process, the influence of difficult samples of different degrees on the model is balanced, the model training time is shortened, and the model precision is improved.

Description

technical field [0001] The invention relates to the technical field of artificial intelligence image processing, in particular to a gradient balance-based multi-category model training method, medium and equipment. Background technique [0002] With the advent of the era of big data and the continuous improvement of computing power, deep learning technology driven by big data has developed rapidly. Deep learning has a wide range of applications in many application fields such as image classification, target detection (typical applications such as face recognition, pedestrian recognition, vehicle recognition, etc.), image segmentation, etc. Behind the applications in these fields are many excellent neural network structures that have emerged in recent years. [0003] It can be said: Mature application model = large amount of data + excellent network structure + suitable training method. Through appropriate training methods and a large amount of data, the network model can a...

AI Technical Summary

Problems solved by technology

For some tasks, too many simple samples will cover the influence of a small number of difficult samples on the model, making it difficult for the model to learn difficult samples; similarly, paying too much attention to difficult samples will also reduce the model's ability to learn simple samples

Therefore, there is currently no training strategy that can simultaneously learn simple samples and difficult samples, which also shows that the current mainstream training methods still need to be further improved.

Images