Norm-based filter pruning method and system for convolutional neural network model

Abstract

The invention discloses a norm-based filter pruning method for a convolutional neural network model, which comprises the following steps of: (1) visualizing all filters of the convolutional neural network model, and clustering the filters into a plurality of clusters according to a visualized output feature map; (2) determining the pruning rate of each cluster according to the average norm of thecluster; (3) judging whether each cluster can be pruned or not according to the model precision loss after each cluster is pruned; and (4) forming a new convolutional neural network model by the pruned structure of each convolutional layer. According to the method, the model is compressed as much as possible under the condition that the precision of the model is ensured, a filter which extracts specific features and is small in norm is reserved while pruning is carried out on the redundant filter, and the functional integrity of the convolutional neural network model is ensured.

Description

Technical field[0001]The present invention belongs to the field of computer technology, and more specifically, relates to a filter pruning method and system based on a norm-based convolutional neural network model.Background technique[0002]With the development of Internet technology and artificial intelligence, Convolution Neural Network (CNN) has shown great advantages in the field of computer vision. However, due to the huge computational cost and memory occupation, it is difficult to operate on platforms with limited resources (such as mobile Client). CNN includes a series of convolutional layers and fully connected layers. Each convolutional layer contains multiple filters. On the premise of ensuring the accuracy of the network model, some filters are pruned to achieve network model compression, which is beneficial to resources Deployment on limited platforms.[0003]Most of the existing pruning methods are sorted by filter norm, and the number of redundant filters is determined b...

AI Technical Summary

Problems solved by technology

[0005] Aiming at the above defects or improvement needs of the prior art, the present invention provides a filter pruning method and system based on a norm-based convolutional neural network model, the purpose of which is to cluster each convolutional layer filter as With multiple clusters, the pruning method of norm sorting is adopted in each cluster, and the pruning method between clusters is used to solve the problem that in the prior art, it is easy to prune a class of filters with small norms and large features, thereby destroying The functional integrity of the convolutional neural network model leads to the technical problem of the loss of precision of the convolutional neural network model and the inability to extract some features

Images