Data processing method and device

Abstract

The invention discloses a data processing method and device, a processor can comprise a cache module and a plurality of logic modules, the processor can obtain image data collected by an image sensorand store the image data in the cache module, and when the image data is processed through a data processing model, for each layer of the data processing model, the data processing module processes the image data. Model parameters of the layer are acquired, the model parameters are loaded to each logic module, an intermediate image obtained from the upper layer of the layer is acquired from the cache module, the intermediate image is segmented to obtain each sub intermediate image, each sub intermediate image is loaded to each logic module, operation on the model parameters and each sub intermediate image are performed through each logic module to obtain a plurality of sub intermediate images, an intermediate image obtained by the layer is acquired, and the intermediate image is stored ina cache module. By means of the method, the image data or the intermediate images obtained in all the layers are stored in the cache, the problem that the data transmission bandwidth is limited is solved, computing power resources of the processor are effectively used, and the effect of reducing time consumption is achieved.

Description

technical field [0001] This specification relates to the technical field of machine learning, and in particular to a data processing method and device. Background technique [0002] At present, in terms of hardware support, the operation of machine learning models can be realized through processors and memories. The processor may include a graphics processing unit (Graphics Processing Unit, GPU), a tensor processing unit (Tensor Processing Unit, TPU), etc., and the memory may include, for example, double data rate synchronous dynamic random access memory (Double Data Rate Synchronous Dynamic Random Access Memory, DDR SDRAM) and other memory components. [0003] Taking target detection on image data and the machine learning model as a Convolutional Neural Networks (CNN) model as an example, the image data, CNN model parameters, and the calculation results of each layer of CNN can be stored in DDR SDRAM. For TPU For the operation of any layer in CNN, first load the CNN model...

AI Technical Summary

Problems solved by technology

[0004] Due to the large amount of data such as image data and CNN calculation results, loading them from DDR SDRAM to TPU requires high bandwidth to support data transmission. Due to the limited transmission bandwidth between DDR SDRAM and TPU, data loading requires It takes a long time, and because the computing power is positively correlated with the demand for bandwidth, when the bandwidth is limited, the computing resources in the TPU cannot be used effectively, and the computing speed of the TPU is low, which causes the TPU to target the image data. Detection takes a long time

Images