- Abstract: Convolutional Neural Network (CNN) has led to great advances in computer vision. Various customized CNN accelerators on embedded FPGA or ASIC platforms have been designed to accelerate CNN and improve energy efficiency. However, the odd-number filter size in existing CNN models prevents hardware accelerators from having optimal efficiency. In this paper, we analyze the influences of filter size on CNN accelerator performance and show that even-number filter size is much more hardware-friendly that can ensure high bandwidth and resource utilization. Experimental results on MNIST and CIFAR-10 demonstrate that hardware-friendly even kernel CNNs can reduce the FLOPs by 1.4x to 2x with comparable accuracy; With same FLOPs, even kernel can have even higher accuracy than odd size kernel.
- Conflicts: huawei.com