Data processing method based on photonic neural network chip and related device or equipment

Abstract

The invention provides a data processing method based on a photonic neural network chip and a related device or equipment, and the method comprises the steps: segmenting a to-be-processed image into aplurality of sub-images with different sizes, carrying out the multilayer convolution processing of each sub-image one by one, and determining the size of a buffer corresponding to each sub-image based on the convolution kernel of each layer and an intermediate result; and determining a target buffer size from the plurality of buffer sizes. According to the scheme, the size of the buffer is set to be larger than or equal to the sum of the occupied sizes of the buffers required by the convolution kernel of each layer and the intermediate result, that is, it is ensured that the convolution kernel and the intermediate data of one sub-image can be stored in the buffer at a time during multi-layer convolution, and repeated data reading is avoided. Since the size of the sub-image is smaller than the size of the image before cutting, the smaller the size of the input multi-layer convolution image is, the higher the convolution calculation efficiency is. According to the scheme, the calculation efficiency is improved while the memory reading bandwidth is reduced by adopting a relatively small buffer size.

Description

technical field [0001] The invention relates to the technical field of data computing, in particular to a data processing method, device, storage medium and electronic equipment based on a photonic neural network chip. Background technique [0002] Convolutional neural network has developed rapidly in the fields of speech recognition and image processing by virtue of its architectural characteristics of local weight sharing. When performing convolution calculations on images, after each layer has completed the convolution operation, the intermediate results of the calculation and the convolution kernel need to be stored in the buffer. [0003] At present, the storage space of the buffer is limited, and all the convolution kernels and intermediate results cannot be stored at one time, and the image needs to be read repeatedly, resulting in low space utilization of the buffer. Therefore, how to provide a data processing method based on the photonic neural network chip, which ...

AI Technical Summary

Problems solved by technology

[0003] At present, the storage space of the buffer is limited, and all the convolution kernels and intermediate results cannot be stored at one time, and the image needs to be read repeatedly, resulting in low space utilization of the buffer.

Images