Convolutional neural network acceleration method and apparatus

Abstract

The invention relates to a convolutional neural network acceleration method and apparatus. The method comprises the steps of reading an input feature graph of a convolutional layer; inputting the input feature graph to a processing array group of the convolutional layer, and according to a first reference pixel vector, a second reference pixel vector, a convolution kernel weight and finished inputchannel data, performing convolution multiplication and addition operation by utilizing a propagate partial multiplier-accumulator to obtain an output result of the processing array group; accordingto the output result of the processing array group, obtaining an output feature graph of the convolutional layer; writing the output feature graph of the final layer of the convolutional layer in an input cache of a full connection layer; executing the multiplication and addition operation by the full connection layer according to the output feature graph of the final layer of the convolutional layer to obtain an output eigenvector of the full connection layer; and outputting the output eigenvector of the final layer of the full connection layer to a fourth block of a first memory. The hardware resources and the power consumption are effectively reduced; and the processing speed of a convolutional neural network is increased.

Description

technical field [0001] The present disclosure relates to the technical field of neural networks, and in particular to a convolutional neural network acceleration method and device. Background technique [0002] Deep learning has shown very good performance in solving many problems such as video recognition, speech recognition and natural language processing. Among the different types of neural networks, convolutional neural networks have been the most intensively studied. The basic structure of the convolutional neural network consists of two layers: one is the feature extraction layer, the input of each neuron is connected to the local area of ​​the previous layer, and the local features are extracted, which is called the convolutional layer; the other is the feature combination The layer is generally composed of a fully connected neural network, which becomes a fully connected layer due to the classification of the image features extracted by the previous feature extracti...

AI Technical Summary

Problems solved by technology

[0003] In view of this, the present disclosure proposes a convolutional neural network acceleration method and device to solve the problem of slow calculation speed of convolutional neural networks

Images