Abstract

Dynamic channel pruning is widely used in model compression to improve the efficiency of neural networks. Although dynamic pruning can remove redundant channels dynamically, the parameters still remain unchanged, which can limit the performance as the response of each neuron changes with different inputs. In this paper, we propose a novel dynamic channel pruning method with adaptive weight learning, which can adaptively adjust both the parameters and widths of the filters simultaneously. Specifically, we design an adaptive-weight convolution module, which can be customized for different inputs under the guidance of global context information to capture representative local patterns and synthesize interested features. At the same time, we utilize a channel importance prediction module to predict the saliency of each channel. Based on the channel saliency, unimportant channels can be removed dynamically to speed up the convolution. These two modules work jointly to achieve a good trade-off between model performance and computational complexity. Experiments on image classification and object detection tasks demonstrate that our method can greatly reduce the computational burden while maintaining the performance, which outperforms state-of-the-art methods.