Data processing method and device and readable storage medium

Abstract

The invention discloses a data processing method and device and a readable storage medium. The data processing method comprises the following steps: receiving input data; processing the data via at least one sub-network model of the neural network model, wherein the neural network model comprises a plurality of basic units and a plurality of sub-network models, each sub-network model comprises atleast part of the basic units in the plurality of basic units, and at least two sub-network models in the plurality of sub-network models share one or more basic units in the plurality of basic units.According to the embodiment of the invention, a plurality of sub-network model structures capable of sharing parameters are designed in a neural network; therefore, a plurality of different sub-network models can be trained at the same time in one training process of the neural network, the precision and speed requirements of different data processing devices can be met, the training and testingprocesses in the actual development process are greatly simplified, and computing resources are saved.

Description

technical field [0001] The present invention relates to the field of data processing, in particular to a data processing method, device and readable storage medium. Background technique [0002] In a series of methods based on deep learning (such as target detection, image classification, object recognition, etc.), the training of the neural network model is a key link related to the accuracy and performance of the model output. [0003] Most of the current model training methods randomly initialize the parameters of the neural network for model training, and then deploy to the corresponding devices. At the time of deployment, computing resources on devices such as mobile devices, desktop devices, DSP devices, GPU devices, etc. have different computing capabilities; and have different preferences for the structure of the neural network model; in addition, the size of the neural network model , different network structures, and even different sub-modules in the same network ...

AI Technical Summary

Problems solved by technology

[0005] Therefore, in the existing technology, in order to adapt to the different computing performance of different hardware devices, it is necessary to initialize and train neural networks of different sizes, so it will take up multiple computing and storage overheads

Images