Channel pruning method based on variational structure optimization network

Abstract

The invention belongs to the technical field of convolutional neural network compression and acceleration, and particularly provides a channel pruning method based on a variational structure optimization network, and the method comprises the steps: compressing a deep convolutional neural network model through a channel pruning technology based on the variational structure optimization network. According to the method, the application limitation of an existing large neural network on resource limitation is considered, a channel pruning technology is adopted to compress an original network, network parameters are compressed as much as possible on the premise that the performance of the original network is not affected, memory occupation of an active layer in the forward propagation process of the network is reduced, and the floating point operation frequency during operation is reduced; therefore, the target of a lightweight network is achieved. By automatically optimizing the network structure, the parameter redundancy of the deep convolutional neural network is effectively reduced, and the operation speed of the deep convolutional neural network is improved, so that the applicationscene of the neural network on edge equipment is expanded.

Description

Technical field[0001]The invention belongs to the technical field of convolutional neural network compression and acceleration, and specifically provides a channel pruning method based on a variational structure optimization network.Background technique[0002]The neural network model is the mathematical expression of the biological neural network learning system. Convolutional neural network (CNN) is one of them. It has demonstrated the most advanced performance in image classification, object detection, image segmentation and other machine vision applications . However, the success of convolutional neural networks often depends on a large number of computing and memory resources. The most advanced models in image classification usually have tens of millions of parameters and require billions of floating point operations to complete the prediction of an image. The huge amount of parameters and calculations make it difficult to deploy in many practical applications of convolutional ne...

AI Technical Summary

Problems solved by technology

[0004] The basic idea of ​​channel pruning is to measure the importance of filters, and then subtract unimportant filters; however, traditional channel pruning is often based on artificially designed methods to measure the importance of filters, which is extremely dependent on expert experience and tends to be one-sided And subjective, and how many unimportant filters are cut is often selected through heuristic or search methods, resulting in cumbersome and inefficient pruning process, and limited compression performance

Images