Class-based filter pruning method

Abstract

The invention relates to a class-based filter pruning method, which comprises the following steps: performing reconstruction based on original VGG-16, adding an activation value generation module, obtaining a new network model, and the activation value generation module being composed of an average pooling layer and two full-connection layers; training the network model added with the activation value generation module, so that the test accuracy of the model on the test set reaches the highest, and obtaining an optimal model; inputting the training set into the optimal model again, and obtaining an activation value of each filter by using an activation value generation module and an input picture; calculating the activation value variances of the filters of all channel importance pruning parts, selecting the filters needing to be reserved according to the variances, and then obtaining the pruning result of the deep network; obtaining a global pruning result of different layers, calculating the number of remaining channels after pruning of each layer, changing the number of channels of each layer of an original network, and retraining the cut original network by using a data set to recover the precision.

Description

technical field [0001] The invention relates to the field of lightweight models in image processing, and is especially suitable for platforms with less computing resources. Background technique [0002] Image classification and detection is a very active research direction in the field of computer vision and machine learning. Image classification and detection are widely used in many fields, including face recognition, garbage classification, behavior recognition, etc. It can be said that object classification and detection have been applied to all aspects of people's daily life. [0003] In the deep learning network, with the deepening of the network depth, the existing convolutional neural network such as VGG [1] 、ResNet [2] 、GoogleNet [3] 、DenseNet [4] The high amount of storage space and computing resource consumption brought about by the amount of calculation and parameters of complex models such as complex models prevents the model from being directly deployed on ...

AI Technical Summary

Problems solved by technology

However, most of the current pruning methods have certain limitations, ignoring that the filters corresponding to different categories of images in the classification network are different, and there is still some redundancy in the retained important filters.

Images