Research On Convolutioanal Neural Network Model Mapping Tool For FPGA

Convolutional neural networks have been applied to many fields such as object detection,image segmentation,and semantic analysis.In many real-world scenarios,there is an increasing trend on performing CNN inference in an embedded-computing context,which calls for the support of low-latency and low-power hardware platforms.Field Programmable Gate Array(FPGA)has been widely used in the research of hardware acceleration of convolutional neural networks due to high performance and low power consumption.However,much FPGA design experience is needed to develop such hardware acceleration.Therefore,this paper proposes a model mapping tool that can automatically generate hardware accelerator code,and at the same time realize automatic software control.In this way,we realize automatic mapping of convolutional neural network to FPGA,and reduce the workload in development.Due to the diversity of the convolutional neural network model,a unified data format is defined in the model mapping tool,and the model file generated by the TensorFlow platform is converted into an intermediate representation.Based on the existing basic hardware architecture,this paper has optimized and designed the hardware template.Then,on one hand,the intermediate representation is used to configure the parameters in the hardware design template and instantiate them;on the other hand,as the input parameter files for the control software,it helps with the memory allocation and runtime control.This allows various convolutional neural networks to be automatically modeled into the FPGA.Meanwhile,in order to verify the practicability of the model mapping tool,it is applied to the real-time target detection system to detect its working condition in the real application scenario.Based on the BT1120 video transmission protocol,image data is collected by the camera device.The image data,through the processing of the de-frame module and the color space conversation module,is used as the input data of CNN model,and the detection result is displayed correctly in real time,which can verify the practicality of the automation tool.In this thesis,several popular target detection models,including SSD(Single Shot Multi-Box Detector),YOLO(You Only Look Once)and ResNet(Residual Network)are mapped to FPGA accelerators through our automation tools.The experimental results show that the automated model mapping tool can ensure the correctness of all calculation results,and only increases the additional consumption time by about 3% compared with the result of manual implementation of the model mapping.In the real-time application,the FPGA acceleration platform uses a 32-bit quantization design.The real-time application,using SSD model,can reach up to 15 frames per second with power consumption of only 4.6W.Therefore,the study of the automation tool in this thesis can map diverse CNN models into FPGA,and realize hardware generation and software control automation,which greatly reduces the difficulty of using FPGAs and reduces programmers' workload and the development cycle.