Platform-aware Efficient Convolutional Neural Network Architecture Search

Computer vision is a kind of applied science which can be widely used in industrial control,self-driving car,environmental monitoring,human-computer interaction,information processing and many other areas.Recently,computer vision developed rapidly with the growth of convolution neural networks.Both in academia and industry,deploying CNNs on embedded devices is a popular and challenging topic.However,CNNs requires huge computing power and high memory read-write throughput.Different from high-end GPUs,embedded platforms have more limitations on hardware resources and bring up many constraints in deployment.What's more,applications run on embedded platforms usually requires real-time latency,a big model with slow infernece speed will become bottleneck of the whole system.Furthermore,different platforms have different computation characters.A highly efficient CNN can become quite slow on another platform.To solve this problem,this paper proposed a novel platform-aware NAS approach and get efficient CNN architectures on given platform.Firstly,this paper introduces the theoretical basis of the neural network,describes the structure and theoretical part of artificial neural network and convolutional neural network,so that readers have a basic concept of computer vision task and convolutional neural network.Then,the design ideas and application techniques of constructing lightweight convolution neural network are introduced,and the lightweight network design suitable for image classification task is extended to semantic segmentation task.Subsequently,this paper focuses on the POP algorithm.In order to search orderly in nearly infinite search space,a new method of structurize search space is proposed and encoded for POP algorithm.Then the design of search strategy is introduced,after that,the partial order hypothesis in the search space is proposed and verified by experiments.According to the partial order relation,the detailed search steps of POP algorithm can be obtained.Similarly,the search space of decoder used in semantic segmentation task can be constructed and the efficient decoder structure can be searched by POP algorithm.At last,many detailed experiments have been done to validate the effectiveness of POP algorithm.On ImageNet dataset and NVIDIA TX 2 platform,DF models proposed in this paper get outstanding leading when compared with human-designed light-weight networks and common networks.Using similar inference time,DF models leads over5% performace,under similar performance,DF models save over 38% inference time.On Cityscapes dataset and NVIDIA 1080 Ti GPU,DF-seg models achieve state-of-the-art trade-off between accuracy and latency and surpasses all human-designed light-weight segmetation models.All experiment shows that POP algorithm can effectively get a series of efficient models on target platform,DF models searched from TX2 platform can achieve high performance on embedded platform and have a wide application prospect.