실험 코드 정리

PPT 실험 파트 구성 순서에 따라 코드 설명

Efficiency 관련 코드

• Model 종류:
  • Vision / HAR / HARNAS
• Metric 종류:
  • Size(MB) / FLOPs(M) / PC:GPU(ms) / Nano:CPU(s) / Nano:GPU(ms) / Smartphone:A31-CPU(ms)
• Device 종류:
  • PC / Nano / Smartphone

Model	Size(MB)	FLOPs(M)	PC:GPU(ms)	Nano:CPU(s)	Nano:GPU(ms)	Smartphone:A31-CPU(ms)
Vision	-	-	-	-	-	-
HAR	-	-	-	-	-	-
HARNAS	-	-	-	-	-	-

✨ 대부분의 실험은 쥬피터 노트북에 정리되어 있음.

1. measure metrics:

>>> python ./ea_harnas/measure_metrics.py --dataset $data --arch EANAS
>>> python ./rl_harnas/measure_metrics.py --dataset $data --arch RLNAS
>>> python ./dnas_harnas/measure_metrics.py --dataset $data --arch OPPA31

2. measure latency:

>>> python ./ea_harnas/measure_latency.py --dataset $data --arch EANAS --num-runs 100 --hardware pc --device gpu
>>> python ./rl_harnas/measure_latency.py --dataset $data --arch RLNAS --num-runs 100 --hardware pc --device gpu
>>> python ./dnas_harnas/measure_latency.py --dataset $data --arch OPPA31 --num-runs 100 --hardware pc --device gpu

3. convert the model(or blocks) for smartphone

>>> python ./convert_models/convert_har_mobile.py --dataset $data --arch EANAS
>>> python ./convert_models/convert_harblock_mobile.py --dataset $data --arch EANAS
>>> python ./convert_models/convert_vision_mobile.py --dataset $data --arch RLNAS
>>> python ./convert_models/convert_visblock_mobile.py --dataset $data --arch RLNAS
>>> python ./convert_models/convert_harnas_mobile.py --dataset $data --arch RLNAS

---

Performance(F1-Score) 관련 코드

• Model 종류:
  • Vision / HAR / HARNAS
• Dataset 종류:
  • UCI-HAR / WISDM / UniMiB-SHAR / OPPORTUNITY / KU-HAR

Model	UCI-HAR	WISDM	UniMiB-SHAR	OPPORTUNITY	KU-HAR
Vision	-	-	-	-	-
HAR	-	-	-	-	-
HARNAS	-	-	-	-	-

1. HAR Model train

   • run_har 디렉토리의 bash 실행

2. Vision Model train

   • run_vis_{dataset} 디렉토리의 bash 실행
   • 각 데이터셋 별로 따로 디렉토리 구성

3. HARNAS Model train

   • run_harnas 디렉토리의 bash 실행

---

HARNAS Model

HARNAS의 model들은 기본적으로 OPPORTUNITY 데이터셋에 최적인 모델 활용.

• RL-based NAS

  • OPPORTUNITY MODEL
  • Dataset이 OPPORTUNITY만 활용됨.
  • Pellatt, Lloyd, and Daniel Roggen. “Fast Deep Neural Architecture Search for Wearable Activity Recognition by Early Prediction of Converged Performance.” In 2021 International Symposium on Wearable Computers, 1–6, 2021.

• EA-based NAS

  • OPPORTUNITY MODEL
  • OPPORTUNITY dataset에 최적인 model을 다른 dataset에 directly 적용했다고 논문들에서 언급하고 있음.
  • Wang, Xiaojuan, Xinlei Wang, Tianqi Lv, Lei Jin, and Mingshu He. “HARNAS: Human Activity Recognition Based on Automatic Neural Architecture Search Using Evolutionary Algorithms.” Sensors 21, no. 20 (October 19, 2021): 6927. https://doi.org/10.3390/s21206927.

• DNAS-based NAS

  • OPPORTUNITY A31 MODEL
  • 위 논문들을 따라 OPPORTUNITY을 활용하여 탐색한 모델 제시.
  • Lim, Won-Seon, Wangduk Seo, Dae-Won Kim, and Jaesung Lee. “Efficient Human Activity Recognition Using Lookup Table-Based Neural Architecture Search for Mobile Devices.” IEEE Access 11 (2023): 71727–38. https://doi.org/10.1109/ACCESS.2023.3294564.