Authors: Jue Li, Feng Cao, Honghong Cheng, Yuhua Qian

Abstract:

Convolutional networks bring the performance of many computer vision tasks to  unprecedented heights, but at the cost of enormous computation load. To reduce this cost, many  model compression tasks have been proposed by eliminating insignificant model structures. For  example, convolution filters with small absolute weights are pruned and then fine-tuned to  restore reasonable accuracy. However, most of these works rely on pre-trained models without  specific analysis of the changes in filters during the training process, resulting in sizable model  retraining costs. Different from previous works, we interpret the change of filter behaviour  during training from the associated angle, and propose a novel filter pruning method utilising the  change rule, which can remove filters with similar functions later in training. According to this  strategy, not only can we achieve model compression without fine-tuning, but we can also find a  novel perspective to interpret the changing behaviour of the filter during training. Moreover, our  approach has been proved to be effective for many advanced CNN architectures.

Keywords： model compress; filter pruning; filter correlation; filter behaviour interpretable

Learning the number of filters in convolutional neural networks.pdf

查看详情

Fri Dec 24 19:35:00 CST 2021