LAUSR

Deep learning has made significant progress in many fields such as image identification, speech recognition and natural language processing, especially in the field of computer vision. The better performance of the neural network often built on deeper, wider network structure, more network parameters and more storage and often computational expensive. As a result, it is hard to deploy neural network to mobile and embedded devices. Therefore, compressing of convolutional neural networks is very necessary and practical. In this paper, we propose a channel pruning algorithm for depth-wise separable convolution units and introduce a new channel selection algorithm based on information gain and a method for quickly recovering network performance after pruning. The proposed method is implemented on MobileNet and validated on several popular datasets. The experimental results show that our method can achieve better experimental results on several image classification datasets, and also achieve good detection results on the PASCAL VOC image detection dataset.