Memory management method, device and storage medium

Abstract

The embodiment of the invention discloses a memory management method, device and storage medium, belonging to the technical field of computers. The method includes: determining at least one branch of the neural network according to the connection relationship of the characteristic units in the neural network; for each branch, according to the size of the output memory required by each characteristic unit on the branch, allocating the first memory and the second memory for the branch , the first memory size is not smaller than the second memory size, and neither the first memory size nor the second memory size is smaller than other memory sizes required by the branch; the first memory and the second memory are used in turn as the input memory and output of the feature unit on the branch Memory. The embodiment of the present invention only allocates two memories for each branch in the neural network, and takes turns as the input memory and output memory of the feature unit, which can not only ensure the normal progress of the calculation, but also realize memory multiplexing, saving the occupied memory, The memory requirement is reduced to ensure that the neural network can be implemented normally on the terminal.

Description

technical field [0001] The embodiments of the present invention relate to the technical field of computers, and in particular to a memory management method, device and storage medium. Background technique [0002] In recent years, deep learning has been widely used in the fields of speech recognition and computer vision, and with the rapid development of deep learning, AI (Artificial Intelligence, artificial intelligence) algorithms are constantly emerging, and rapidly change the development of technology. direction and people's lives. [0003] Deep learning is inseparable from the implementation of neural networks. Traditional neural networks are usually implemented on the server side. However, with the continuous development of AI algorithms, the way to implement neural networks on the server side can no longer meet people's growing business needs. Propose a way to implement neural networks in the terminal. [0004] However, the neural network includes multiple network l...

AI Technical Summary

Problems solved by technology

[0003] Deep learning is inseparable from the implementation of neural networks. Traditional neural networks are usually implemented on the server side. However, with the continuous development of AI algorithms, the way to implement neural networks on the server side can no longer meet people's growing business needs. Propose a way to realize the neural network in the terminal

[0004] However, the neural network includes multiple network layers, and each network layer includes multiple feature units. Each feature unit needs to occupy a piece of memory when the output data is calculated, resulting in the need to occupy a large amount of memory when implementing the neural network.

Compared with the server side, the terminal has limited memory and cannot meet the memory requirements of the neural network, making it difficult to realize the neural network in the terminal.

Images