Research And Optimization Of FairMOT Model Based On Ascend AI Processor

With the development of deep learning,it has a wider range of application in life.The excellent effect presented by the deep learning model need the high-performance computing equipment.However,the United States has now stepped up the export of high-performance computing equipment to China,and it is extremely urgent to migrate the model to domestic computing equipment.This paper takes FairMOT model as the starting point,and research migration based on Huawei Ascend AI processor.However,there are some problems.The software stack of Ascend AI processor is not perfect,and the model cannot run directly on the Ascend AI processor.No one has migrated the Object Tracking model to the Ascend AI processor before,and the migration process scheme is not perfect.This paper makes the following work for the above problems:(1)By analyzing the characteristics of FairMOT model in training and inference,a migration scheme is proposed to locate most missing operators in the early stage to avoid repeated troubleshooting.(2)Based on the preliminary analysis and preparation of the migration work,Cos operator is developed through DSL development mode,Matrix Power operator is developed through TIK development mode,Top K operator is developed through AI CPU development mode,and the performance of the developed operator is optimized based on the hardware characteristics of Ascend processor.(3)Complete the preliminary operation on Ascend processor by integrating the developed operator with the upper framework.Through the mixed precision training based on Apex library,DDP mode distributed training,making full use of the processor performance.By binding the CPU core and avoiding dynamic tensors,the training speed on Ascend servers can be improved.Compared to the version that haven’t been optimized,the speed has increased by 32%.(4)Transform the model to obtain a special model structure that can run on the Ascend server,and write the pre and post processing plugins through Mind SDK,making full use of the CANN architecture efficiency to complete the inference migration of the model on the Ascend processor.Through optimizing measures such as operator fusion and adding block reasoning speed,the inference speed has been improved by 46%compared to the unoptimized version.