Efficient Computation of Depthwise Separable Convolution in MoblieNet Deep Neural Network Models

MobileNets are light-weight deep neural network (DNN) models with fewer parameters and less computation compared with some popular models such as AlexNet and VGG. MobileNets adopt depthwise separable convolution (DSC) composed of 1x1 pointwise convolution (PWC) and depthwise convolution (DWC) to reduce model complexity. In this paper, we present a DNN accelerator design which can perform the DSC more efficiently by combining the PWC and DWC with fewer external memory accesses. Experimental results show that the proposed design can reduce more than 50% DRAM accesses with increased speed.