Hardware Implementation and Quantization of Tiny-Yolo-v2 using Open CL

The trend of increasingly model size in Deep Neural Network (DNN) algorithms boost the performance of visual recognition tasks. These gains in performance have come at a cost of increase in computational complexity and memory bandwidth. Recent studies have explored the fixed-point implementation of DNN algorithms such as AlexNet and VGG on Field Programmable Gate Array (FPGA) to facilitate the potential of deployment on embedded system. However, there are still lacking research on DNN object detection algorithms on FPGA. Consequently, we propose the implementation of Tiny-Yolo-v2 on Cyclone V PCIe FPGA board using the High-Level Synthesis Tool: Intel FPGA Software Development Kit (SDK) for OpenCL. In this work, a systematic approach is proposed to convert the floating point Tiny-Yolo-v2 algorithms into 8-bit fixed-point. Our experiments show that the 8-bit fixed-point Tiny-Yolo-v2 have significantly reduce the hardware consumption with only 0.3% loss in accuracy. Finally, our implementation achieves peak performance of 31.34 Giga Operation per Second (GOPS) and comparable performance density of 0.28GOPs/DSP to prior works under 120MHz working frequency.