목적

• 기존에 만들었던 모델 보완
• 기존 모델이 shifting window length는 9 frames을 사용하지만, snr을 연산할 때 25 frames을 사용하는 괴리가 있었음
• snr을 9 frames으로 연산하도록 수정하여 모델을 다시 훈련시킴
• 파일명: S03_features_v2 & S04_labeling_v2

모델 재훈련

기존

Train Loss: 0.0843, Train Accuracy: 0.9791
Test Loss: 0.1109, Tset Accuracy: 0.9597

• Voting 조건 Grid Search

  Best Configuration:
  threshold                      0.55
  voting_type             consecutive
  voting_size                      10
  train_accuracy                100.0
  test_accuracy                 100.0
  train_lead_time_mean     524.222222
  train_lead_time_std       107.92979
  test_lead_time_mean      540.952381
  test_lead_time_std        90.390937
  

  

변경

Train Loss: 0.0929, Train Accuracy: 0.9735
Test Loss: 0.1073, Tset Accuracy: 0.9646

• Voting 조건 Grid Search

  • voting까지 15 counts 이내에 이루어지도록

  Best Configuration:
  threshold                      0.9
  voting_type               majority
  voting_size                     10
  train_accuracy           99.444444
  test_accuracy                100.0
  train_lead_time_mean         500.0
  train_lead_time_std     123.178869
  test_lead_time_mean     515.238095
  test_lead_time_std      124.809151
  

  

성능 확인

• consecutive voting x 4 & threshold 0.9

  # Training
  Sensitivity: 98.89 % (89/90)
  Specificity: 100.00 % (90/90)
  Accuracy: 99.44 % (179/180)
  Lead Time: 587.19 ± 140.60 ms
  

  # Test
  Sensitivity: 100.00 % (21/21)
  Specificity: 100.00 % (21/21)
  Accuracy: 100.00 % (42/42)
  Lead Time: 597.14 ± 145.05 ms
  

TF Lite 변환

성능 체크

TF Lite Micro 변환