A progressive model compression method for complex network structure

Abstract

The application belongs to the field of lightweight convolutional neural network, and relates to a progressive model compression method for complex network structure. The application aims to solve the problems that the existing lightweight network design method is difficult to be applied to a complex basic model, and the existing lightweight network is not suitable for a scene with high requirements for detection speed and detection accuracy. The process is as follows: step 1, performing filter pruning and channel number alignment on a target model for image recognition; step 2, giving a loss function of the target model after filter pruning and channel number alignment in step 1; inputting a feature map into the target model after filter pruning and channel number alignment in step 1, and each convolution layer l of the target model outputs y l , using a back propagation algorithm to calculate the gradient T l of the loss function with respect to y l , and replacing the convolution operation in the convolution layer with a lightweight convolution operation according to the gradient.

Images