Fast Depthwise Separable Convolution for Embedded Systems

摘要

Convolutional neural networks (CNNs) have achieved outstanding performance in many applications. However, as the total number of layers has increased and the model structure has become compound, the computational cost comes into question. The large models cannot operate in embedded or mobile environments where hardware resources are quite limited. To overcome these problems, there have been several attempts like reducing the depth of networks, pruning, quantization or low rank approximation. Depthwise separable convolution (DSC) was proposed to reduce computation especially in convolutional layers by separating one convolution into a spatial convolution and a point-wise convolution. In this paper, we apply DSC to the YOLO network for object detection and propose a faster version of DSC, FastDSC by replacing the pointwise convolution with general matrix multiplication. Experiments on the NVIDIA Jetson TX2 board show that FastDSC speeds up DSC for object detection.