Convolutional neural network feature map data compression method and device

Abstract

The invention discloses a convolutional neural network feature map data compression method and device. The method comprises a feature map channel reconstruction stage and a zero-value bit map coding compression stage. In the feature map channel reconstruction stage, feature map channel dimensions are reconstructed by using one-dimensional discrete cosine transform, high-frequency information filtering is realized by using a frequency domain filter, a sparse feature map of which the sparseness is higher than that of an original feature map is obtained, and then the sparse feature map is compressed in the zero-value bit map coding compression stage, and final compressed data of the original feature map is obtained. According to the method, channel redundancy of the convolutional neural network is utilized, channel groups with certain frequency domain features are reconstructed together, the sparseness of the data to be transmitted is improved, the compressed potential is further improved, then zero-value bit map coding compression is used for compressing the sparse feature map, and the compression rate is improved.

Description

technical field [0001] The present application relates to the technical field of convolutional neural networks, in particular to a data compression method and device for feature maps of convolutional neural networks. Background technique [0002] In the field of machine vision, the network architecture with convolutional neural network as feature extractor has excellent accuracy and high computational efficiency. The convolutional neural network is composed of several or even hundreds of convolution operation layers stacked and connected. Based on the two-dimensional plane convolution calculation, the feature extraction of the image is performed to obtain the feature map data. [0003] In practical applications, the convolutional neural network is usually deployed to a hardware accelerator or other low-power devices. After the feature map data is generated by the activation function in the convolutional neural network, it is transmitted to the off-chip memory of the hardware...

AI Technical Summary

Problems solved by technology

[0005] Zero-value bitmap coding has the advantage of high compression rate, but it also makes the hardware complexity higher

In addition, a small number of hardware accelerators provide the implementation of activation functions such as LeakyReLU. The feature map data generated by this type of activation function does not have a high degree of sparsity, and is not suitable for data compression using the aforementioned zero-value bitmap encoding method. At present, there is no applicable, particularly hardware-efficient compression method. Some available complex encoding compression schemes such as LZW encoding and Huffman encoding are schemes that have a large hardware overhead and are not suitable for hardware implementation.

Images