Fixed-point operation method and device for convolutional neural network, equipment and storage medium

Abstract

The embodiment of the invention discloses a fixed-point operation method and device for a convolutional neural network, equipment and a storage medium, the convolutional neural network comprises a convolutional layer, the method comprises the steps of receiving an input activation value of the convolutional layer through an input channel, and the input channel has a corresponding weight; Performing fixed-point operation on the input activation value to obtain a first characteristic value; respectively writing the first characteristic value and the weight into registers of a plurality of register groups; and for the plurality of register groups, respectively carrying out multiplication and addition operation according to the first characteristic values and the weights in the registers to obtain a plurality of second characteristic values. Due to the fact that a plurality of registers are usually provided in the processor, accumulation operation can be carried out in the plurality of registers in a dispersed mode, namely grouping accumulation is carried out, the number of evenly-shared multiplication and addition operations is reduced, the overflow risk is reduced, the processing efficiency of application operation instructions is improved, the overall throughput capacity is increased, meanwhile, accuracy is kept, and the application range is guaranteed.

Description

technical field [0001] Embodiments of the present invention relate to deep learning technology, and in particular to a fixed-point computing method, device, device, and storage medium of a convolutional neural network. Background technique [0002] In recent years, deep learning has been widely used in vision and other fields. Among them, a series of algorithms centered on CNN (Convolutional Neural Network) have better results in image classification, object detection, and pixel-level segmentation. [0003] However, the convolutional neural network CNN has a large amount of calculations, and more than 90% of its calculation load is concentrated in the convolution operation. The usual implementation will convert the convolution layer into the multiplication operation of two matrices, and the matrix multiplication The core part is as follows: [0004] [0005] Among them, w is the weight, a is the input activation value, O is the output activation value, and N is the commo...

AI Technical Summary

Problems solved by technology

[0017] However, the accuracy of this method is low, the scope of application is small, and it is only suitable for simple image classification tasks

Images