A method and device for constructing a neural network model

Abstract

The invention discloses a method for constructing a neural network model, comprising the steps of: S1 constructing a strategy network and an image test set; S2 constructing a plurality of random code arrays, and inputting them into the strategy network respectively to obtain multiple initial codes; S3 A plurality of neural network models are obtained after processing a plurality of initial codes; S4 inputs the image test set into a plurality of neural network models to obtain multiple rewards, loss values ​​of a plurality of neural network models and a plurality of actual classification results; S5 According to the theoretical classification of the image test set, it is judged whether at least one of the multiple actual classification results meets the preset condition; S6 responds to not satisfying the preset condition, using each initial code and the reward calculation strategy network obtained by its corresponding neural network model The loss value; S7 updates the policy network according to the loss value; S8 repeats steps S2‑S7 until at least one actual classification result meets the preset condition in step S5. The above method can greatly speed up the optimization speed.

Description

technical field [0001] The present invention relates to the field of image classification, more specifically, to a method and device for constructing a neural network model. Background technique [0002] The neural network model is a model structure that can be stacked arbitrarily. The basic components include FC (full connection layer), Convolution (convolution layer), Polling (pooling layer), Activation (activation function), etc. The latter component is the previous one. The output of the component is used as the input, and different component connection methods and hyperparameter configuration methods have different effects in different application scenarios. Neural Architecture Search (NAS), the goal is to search for an optimal neural network model from a bunch of neural network components. Among them, common search methods include random search, Bayesian optimization, evolutionary algorithm, reinforcement learning, gradient-based algorithm, etc. [0003] Zoph et al. ...

AI Technical Summary

Problems solved by technology

However, these algorithms all use the method of manually setting the network architecture, which leads to challenges in the flexibility of the architecture.

Moreover, these algorithms only consider the calculation accuracy of the model, and only one model is used to update the network parameters at a time, resulting in a slow search for the optimal model.

Images