Reconfigurable hardware generation for tensor flow models of CNN algorithms on a heterogeneous acceleration platform

Abstract

Convolutional Neural Networks (CNNs) have been used to improve the state-of-art in many fields such as object detection, image classification and segmentation. With their high computation and storage complexity, CNNs are good candidates for hardware acceleration with FPGA (Field Programmable Gate Array) technology. However, much FPGA design experience is needed to develop such hardware acceleration. This paper proposes a novel tool for design automation of FPGA-based CNN accelerator to reduce the development effort. Based on the Rainman hardware architecture and parameterized FPGA modules from Corerain Technology, we introduce a design tool to allow application developers to implement their specified CNN models into FPGA. Our tool supports model files generated by TensorFlow and produces the required control flow and data layout to simplify the procedure of mapping diverse CNN models into FPGA technology. A real-time face-detection design based on the SSD algorithm is adopted to evaluate the proposed approach. This design, using 16-bit quantization, can support up to 15 frames per second for 256*256*3 images, with power consumption of only 4.6 W.