FPGA-based data processing method and device, and storage medium

Abstract

The invention discloses a FPGA-based data processing method, which is used for realizing data processing of a neural network with multiple hidden layers, and comprises the following steps: S1, processing data by utilizing a first hidden layer, and storing the processed data in a memory; s2, dividing the data processed by the upper hidden layer in the memory into a plurality of parts, and inputtingthe parts into the lower hidden layer in sequence; s3, the next hidden layer processes a part of the input data in the plurality of parts and outputs the processed data to a memory at the same time,and the steps are repeated until the processing of the plurality of parts is completed so as to obtain the data which are stored in the memory and processed by the next hidden layer; s4, repeating thestep S2; And S3, obtaining the data which is stored in the memory and processed by the last hidden layer. According to the FPGA-based data processing method provided by the invention, the data to beprocessed can be input into the corresponding hidden layer in multiple times, so that the calculation efficiency is improved.

Description

technical field [0001] The present invention relates to the field of data, and more specifically, to an FPGA-based data processing method, device and readable storage medium. Background technique [0002] CNN (Convolutional Neural Network) is a kind of artificial neural network, which is widely used in image classification, target recognition, behavior recognition, speech recognition, natural language processing and document classification and other fields. [0003] The structure of CNN generally consists of an input layer, several hidden layers and an output layer. The input layer is used for multi-dimensional input data (such as color pictures), the output layer is used to output recognition results, and the hidden layer is used for neural network calculations. Neural network calculation steps include convolution, pooling, Batch Norm, full connection calculation and activation function, etc. The input of the first hidden layer is the multi-dimensional input data input th...

AI Technical Summary

Problems solved by technology

[0005] The traditional FPGA-based CNN data processing method is to store the convolution kernel and hidden layer feature map data in the on-chip memory. This storage method maximizes the read and write access rate of the on-chip memory, but is limited by the on-chip Once the data volume of a certain intermediate layer feature map of a certain CNN structure exceeds the limit of the on-chip FPGA storage space, the CNN structure will not be able to be implemented on this FPGA.

Images