Design And Implementation Of The Convolutional Neural Network Based On FPGA

CPUs and GPUs require high performances in accelerating in deep learning.FPGA based on hardware devices is a more reasonable choice when thinking of balancing among energy efficiency,space saving and speed.It is feasible to accelerate the application of FPGA in deep learning,because computations can be parallel in layers in Convolutional Neural Network.In the paper,the thought is designed for a network model and completes the forward propagation progress in CNNs to save overall time in the network training.The system analyzes layers,and design layers constructions and connections between layers based on hardware to achieve efficient and high-speed procession.In this work,we present Policy network model which is adopted by AlphaGo.Policy network concludes convolution layer,activation functions,Scale layer and Loss layer and the purpose is to predict and select moves in the game of Go.The number of convolution layers in the model is up to 13.We purpose to divide these layers into three parts and present parallel acceleration module in layers.We complete the system on VCU118 which provides enough resources.Compared with GPU and CPU,we test the results in the same process of forward propagation.As a result,the design can achieve a speed-up computation progress.Therefore,the system has achieved the expected function of entire forward propagation progress CNN model on FPGA,which competes the expected acceleration performance.