Image recognition method based on weight pruning quantization

Abstract

The invention provides an image recognition method based on weight pruning and quantization, and relates to the field of computer vision. The method comprises the following steps: removing a part with a smaller absolute value in a convolutional neural network model of a part of the part, so that the network is sparse; retraining the convolutional neural network model until the model converges, and repeating the above two steps; pruning 10% of the parameters each time, and retraining the pruned convolutional neural network model until 80% of the parameters are cut off; quantizing the weight of the 50% convolutional neural network model by using an 8-bit fixed point, retraining the convolutional neural network model until the model converges, not updating the quantized weight, updating the non-quantized weight, and enabling that the cut weight also participates in updating, and the important pruned weight is reactivated; and repeating the previous steps, and continuously quantizing 10% of parameters. Redundant neurons or weights are cut off, so that the parameters of the model can be effectively reduced, and the calculation amount and the storage space of the model are reduced.

Description

technical field [0001] The invention belongs to the technical field of computer vision and artificial intelligence processing, in particular to an image recognition method based on weight pruning and quantization. Background technique [0002] With the continuous development and upgrading of computer hardware, especially the rapid development of image processors in parallel computing, researchers have obtained much higher computing power than before for the study of convolutional neural networks. Due to the shortcomings of convolutional neural network, such as large volume, many parameters, and large amount of calculation, its deployment and use on mobile devices or low-power devices are limited. At the same time, the complex convolutional neural network occupies more memory bandwidth and increases more energy consumption during inference, making the application experience on the terminal poor. Contents of the invention [0003] The purpose of the present invention is to ...

AI Technical Summary

Problems solved by technology

Due to the shortcomings of convolutional neural network, such as large volume, many parameters, and large amount of calculation, its deployment and use on mobile devices or low-power devices are limited.

At the same time, the complex convolutional neural network occupies more memory bandwidth and increases more energy consumption during reasoning, making the application experience on the end poor.

Images