Method and apparatus for adapting feature data in a convolutional neural network

Abstract

A method and apparatus for adapting feature data in a convolutional neural network are disclosed. The method includes: selecting a plurality of consecutive layers; determining the expected number of sub-data blocks of the output feature data of the last layer and the layout position, width and height of each sub-data block; determining each sub-data of the input feature data of each layer The layout position, width and height of the block; determine the actual position of each sub-data block of the input feature data of the first layer; and according to the actual position, width and height of each sub-data block of the input feature data of the first layer, obtain the most The expected number of input subdata of the input feature data of the first layer. Through this method, the convolution operation of feature data of any size can be efficiently realized while effectively reducing a large amount of data transfer between the external memory and the processor.

Description

Technical field [0001] The present application generally relates to the technical field of artificial neural networks, and specifically relates to methods and devices for adapting feature data in convolutional neural networks. Background technique [0002] Deep learning technology based on convolutional neural networks has been widely used in different fields such as image recognition, video analysis, natural language processing, and driving assistance. It is expected that hardware such as a general-purpose central processing unit (CPU), graphics processing unit (GPU), or a dedicated accelerator can be used to efficiently perform operations in a convolutional neural network. [0003] The access speed of data is an important factor affecting the efficiency of computing. To this end, a high-speed memory such as a cache or on-chip memory may be provided for the processor (for example, a CPU, a GPU, or a dedicated accelerator) to cache at least a part of the data. However, due to lim...

AI Technical Summary

Problems solved by technology

However, due to constraints such as cost, the capacity of high-speed memory is usually limited, so it may not be possible to cache all the data (e.g., feature data) of multiple layers of the convolutional neural network, resulting in high-speed memory relative to the access speed. Between other lower memories (for example, a random access memory coupled to a processor through a bus, or a storage such as a hard disk coupled to a processor or a computing device containing the processor through an interface or data lines) massive data transfer

Images