http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
신상준(Sangjune Shin),신동균(Dongkun Shin) 대한전자공학회 2023 대한전자공학회 학술대회 Vol.2023 No.6
This paper addresses the limitations of realtime application of object detection techniques in embedded computing environments. To overcome these limitations, compression methods such as pruning and quantization are being investigated. In this study, various quantization techniques, including post-training quantization, partial quantization, and quantization-aware training, are applied to the YOLO v3 model to evaluate and analyze their effects in embedded system environments. Experiments were conducted using the YOLO v3 model and the PASCAL VOC dataset. The experimental results show that the quantized models can reduce model size and inference time but result in accuracy loss. Therefore, this research contributes to improving object detection performance in embedded systems by proposing and analyzing various quantization techniques for lightweighting object detection technology in embedded environments.
Vector engine 기반 NPU를 위한 grouped pattern-wise pruning
서수희(Soohee Seo),이하윤(Hayun Lee, Sangho Lee),이상호(Dongkun Shin),신동균(Dongkun S) 대한전자공학회 2023 대한전자공학회 학술대회 Vol.2023 No.6
Among the recent pruning techniques, it is conducted in various ways such as block-wise and channel-wise, and the accuracy is low because it is coarser than the relatively dense model. In addition, there is a disadvantage that hardware must be designed accordingly. To solve this problem, we propose an optimized pruning technique for VTA accelerator GEMM Core. Accuracy was 5.73% to 23.83% higher than block-wise pruned model. It is expected that there will be scalability that can be applied to various vector engine-base NPUs using this technique.