Full connection layer compression method and device, electronic equipment, accelerator and storage medium

Abstract

The invention discloses a full connection layer compression method and device, an accelerator, electronic equipment and a storage medium, and is applied to the technical field of neural networks. Comprising the steps of dividing all layers in a full connection layer into alternate first layers and second layers, compressing a group of first layers and second layers into a binding layer, where ping-pong storage does not exist between the first layers and the second layers; compressing two adjacent layers into one binding layer, so that the intermediate result storage overhead can be effectively reduced. In addition, as the full connection layer is compressed, the flow series of the full connection layer is also changed into half of the flow series before compression, so that the output delay is greatly shortened, and the recognition real-time performance of the accelerator is improved.

Description

technical field [0001] The present application relates to the technical field of neural networks, and in particular to a fully-connected layer compression method, device, accelerator, electronic equipment, and storage medium. Background technique [0002] Convolutional Neural Networks (CNN, Convolutional Neural Networks) is one of the representative algorithms of deep learning. Because of its excellent performance in the field of artificial intelligence, CNN has been widely concerned and applied to high-tech applications such as image classification, speech recognition, face recognition, automatic driving, and medical imaging. At present, excellent CNN structures such as AlexNet, VGG, and ResNet have emerged. With the continuous development of CNN, the network structure becomes increasingly complex and the number of parameters explodes, which brings challenges to the design of CNN hardware accelerators. [0003] The fully-connected layer structure of batch operations can r...

AI Technical Summary

Problems solved by technology

At the same time, the data preparation time Tbatch of the batch operation is NumB times of the ordinary pipeline time Tpipe, so the batch operation will significantly increase the output delay (IRL, Input Response Latency)

Images