Model Channel Pruning Method Based on Squeeze-and-Excitation Mechanism and Upper Quartile Truncation
Zening Ding, Jianyu Zhao, Jiaqi Sun · 2021
As the application of neural networks becomes more profound and more widespread, it is common to deepen the network or increase the number of parameters to improve its performance. Nevertheless, the accompanying increase in computational complexity makes the model difficult to implement on resource-limited mobile devices. Pruning connections are one of the main methods used for deep network compression. The existing pruning techniques suffer from disadvantages such as generating additional energy consumption, short practical estimation of channel importance, and leading the network to fall into overfitting. In this paper, we propose a channel pruning method based on the SE mechanism and upper quartile truncation, which uses the SE mechanism to extract the channel feature importance and upper quartile points as truncation factors to accelerate the model and reduce its computational effort, and the effectiveness of the method was demonstrated by experiments.