Hyper-parameter determination method, device and equipment

Abstract

The embodiment of the invention discloses a hyper-parameter determination method, device and equipment. The method comprises the steps of taking a value of each hyper-parameter of a classification model to form a preset number of numerical value combinations; taking a value of a hyper-parameter in the numerical value combination as a value of a hyper-parameter of a classification model, obtaininga plurality of evaluation indexes of the classification model, obtaining a corresponding weight according to a preset weight rule, and finally determining an optimal hyper-parameter of the classification model from the mapping relation by establishing a mapping relation between the numerical value combination and the weight.

Description

technical field [0001] This specification relates to the field of computer technology, in particular to a method, device and equipment for determining hyperparameters. Background technique [0002] The hyperparameters of the classification model are the external parameters of the model that need to be set before the model is learned, and the hyperparameters cannot be obtained through model learning like the internal parameters of the model, and the hyperparameters are not only a simple value, but also control the The behavior of model training also affects the classification performance of the model to a large extent, so whether the preset hyperparameters are appropriate will not only affect the learning performance of the model, but also affect the classification effect of the model. [0003] At present, hyperparameters are often tuned manually, that is, the possible optimal hyperparameter value is selected according to human experience, and then the selected optimal hyperp...

AI Technical Summary

Problems solved by technology

[0002] The hyperparameters of the classification model are the external parameters of the model that need to be set before the model is learned, and the hyperparameters cannot be obtained through model learning like the internal parameters of the model, and the hyperparameters are not only a simple value, but also control the The behavior of model training also affects the classification performance of the model to a large extent, so whether the preset hyperparameters are appropriate will not only affect the learning performance of the model, but also affect the classification effect of the model

[0003] At present, hyperparameters are often tuned manually, that is, the possible optimal hyperparameter value is selected according to human experience, and then the selected optimal hyperparameter is used as the hyperparameter of the model, and the model is adjusted through the data set. The evaluation index of the model is obtained through training, and then the hyperparameters, training model, and evaluation index of the model are obtained over and over again, and finally the optimal hyperparameters of the model are determined from these hyperparameters according to the evaluation index, so that based on the determined optimal hyperparameters parameters, although it is possible to get a model with better performance, but due to the dependence on human experience, deviations are prone to occur, and it is difficult to obtain the real optimal hyperparameters, and manual tuning is time-consuming and laborious, and it is impossible to simultaneously adjust the hyperparameters in big data applications. Tuning the hyperparameters of the model when applied to multiple tasks

Although there are also automatic tuning methods, such as web search, random search, Bayesian optimization methods, etc., these methods require a lot of computing power and still take a long time, so in big data applications, there is no guarantee that the model will Can determine optimal hyperparameters when applied to multiple tasks

Images