An Automatic Conversion Tool for Caffe Neural Network Configuration oriented to OpenCL-based FPGA Platforms

Recently, the research of image classification and recognition in deep learning area has reached an unprecedented break with the highest accuracy rating that seemed most nearly human. However, the pervasive application of deep learning is critically limited by the issues of efficiency and power consumption. Fortunately, FPGA (Field Programmable Gate Array) provides an effective solution for the on-line inference of various deep learning applications. Nonetheless, the general program designed for CPU and GPU application cannot be directly applied in the FPGA heterogeneous computing platform. In this paper, we focus on the migration of neural network configurations from CPU/GPU-based computing platforms to FPGA-based computing platforms under the framework of Caffe. Our contribution presents an automatic conversion tool to apply the neural network configuration in Caffe model to FPGA-based heterogeneous acceleration platforms. This method takes full advantage of the dictionary and template functions of Python program language. Firstly, we should construct the mustache template for head files according to the neural network model. After finishing the template construction, Python program reads and analyzes the neural network configuration in the prototxt file supported by Caffe framework. Next, it fills in all the corresponding positions of the template with the analytic parameters for neural network configuration. Furthermore, it generates a C++head file that configures the neural network for OpenCL program. This tool helps to achieve the migration of Caffe model from CPU/GPU platform to OpenCL-based FPGA heterogeneous acceleration platforms. Finally, we illustrate that our method promotes the implement of FPGA applications in artificial intelligence field with a high efficiency.