Channel-Wise Pruning via Learn Gates&BN for Object Detection Tasks

Network pruning is an important research area aimed at addressing the high computational costs of deep neural networks. Previous studies [1] have indicated that it is not necessary to follow the conventional pruning process of training a large, redundant network, but rather, more diverse and higher-performing potential models can be directly pruned from randomly initialized weights. However, our experimental results on the MS COCO 2017 dataset [2] demonstrate that this approach is not applicable when compressing object detection models. We believe this outcome is associated with the complex network structures involved in object detection, making it relatively challenging to explore pruned architectures from random weights. To address this issue, we improved the existing Learn gates method and incorporated Batch normalization [3] to jointly learn channel importance. This enhances the learning capability of channel importance in a shorter time frame and facilitates the exploration of suitable pruned network structures within pre-trained weights. When applying our network pruning method to object detection models YOLOv3 [4] and YOLOv4 [5], our approach achieves higher accuracy with only a brief period of network structure learning.