Modeling - [Try] ResNet 층별 성능 비교

[UGeunJi]

ResNet은 18, 34, 50, 101, 152층이 있다. Renewed_Dataset (6 class, 1200 images)으로 층별 성능 검사를 실시했다.

ResNet18

Epoch 1/10
train | Loss: 1.5817 Acc: 0.4396
test | Loss: 1.6308 Acc: 0.4417
Epoch 2/10
train | Loss: 1.2850 Acc: 0.5406
test | Loss: 1.4312 Acc: 0.5500
Epoch 3/10
train | Loss: 1.0194 Acc: 0.6406
test | Loss: 0.7170 Acc: 0.7083
Epoch 4/10
train | Loss: 0.8189 Acc: 0.7146
test | Loss: 0.7164 Acc: 0.6833
Epoch 5/10
train | Loss: 0.6407 Acc: 0.7792
test | Loss: 0.6342 Acc: 0.7750
Epoch 6/10
train | Loss: 0.5825 Acc: 0.8073
test | Loss: 0.3426 Acc: 0.8583
Epoch 7/10
train | Loss: 0.4325 Acc: 0.8490
test | Loss: 0.3481 Acc: 0.8250
Epoch 8/10
train | Loss: 0.3865 Acc: 0.8604
test | Loss: 0.2532 Acc: 0.8583
Epoch 9/10
train | Loss: 0.3185 Acc: 0.8875
test | Loss: 0.2188 Acc: 0.8833
Epoch 10/10
train | Loss: 0.2582 Acc: 0.9010
test | Loss: 0.2994 Acc: 0.9167
<__main__.TrainModel at 0x7f1475788fd0>

test | Loss: 0.2766 Acc: 0.8833

ResNet34

Epoch 1/10
train | Loss: 2.7944 Acc: 0.1500
test | Loss: 2.3708 Acc: 0.1750
Epoch 2/10
train | Loss: 2.1064 Acc: 0.1583
test | Loss: 2.2992 Acc: 0.2667
Epoch 3/10
train | Loss: 2.0260 Acc: 0.2146
test | Loss: 2.2444 Acc: 0.2333
Epoch 4/10
train | Loss: 1.9805 Acc: 0.2167
test | Loss: 1.8449 Acc: 0.2917
Epoch 5/10
train | Loss: 1.8842 Acc: 0.2531
test | Loss: 1.7992 Acc: 0.2667
Epoch 6/10
train | Loss: 1.8348 Acc: 0.2760
test | Loss: 1.6015 Acc: 0.3917
Epoch 7/10
train | Loss: 1.7195 Acc: 0.3167
test | Loss: 1.6189 Acc: 0.4167
Epoch 8/10
train | Loss: 1.5831 Acc: 0.4229
test | Loss: 1.8296 Acc: 0.3167
Epoch 9/10
train | Loss: 1.4798 Acc: 0.4677
test | Loss: 1.2517 Acc: 0.4917
Epoch 10/10
train | Loss: 1.2456 Acc: 0.5458
test | Loss: 1.6936 Acc: 0.4750
<__main__.TrainModel at 0x7f13af7ffc70>

test | Loss: 1.2378 Acc: 0.5083

ResNet50

Epoch 1/10
train | Loss: 4.0782 Acc: 0.1146
test | Loss: 26.3824 Acc: 0.1750
Epoch 2/10
train | Loss: 2.4253 Acc: 0.1562
test | Loss: 3.1925 Acc: 0.1500
Epoch 3/10
train | Loss: 2.2352 Acc: 0.1771
test | Loss: 4.2339 Acc: 0.2333
Epoch 4/10
train | Loss: 2.1786 Acc: 0.2083
test | Loss: 3.8340 Acc: 0.2417
Epoch 5/10
train | Loss: 2.0408 Acc: 0.2406
test | Loss: 3.2647 Acc: 0.2667
Epoch 6/10
train | Loss: 2.0722 Acc: 0.2281
test | Loss: 3.1297 Acc: 0.3000
Epoch 7/10
train | Loss: 1.9709 Acc: 0.2448
test | Loss: 2.8438 Acc: 0.3000
Epoch 8/10
train | Loss: 1.9322 Acc: 0.2750
test | Loss: 2.1233 Acc: 0.3167
Epoch 9/10
train | Loss: 1.9520 Acc: 0.2729
test | Loss: 4.5353 Acc: 0.2750
Epoch 10/10
train | Loss: 1.8970 Acc: 0.2698
test | Loss: 1.9957 Acc: 0.4500
<__main__.TrainModel at 0x7f146d4af970>

test | Loss: 1.8604 Acc: 0.4000

ResNet101

Epoch 1/10
train | Loss: 3.6438 Acc: 0.1281
test | Loss: 2.8691 Acc: 0.1167
Epoch 2/10
train | Loss: 2.3420 Acc: 0.1365
test | Loss: 2.2486 Acc: 0.1417
Epoch 3/10
train | Loss: 2.2081 Acc: 0.1427
test | Loss: 4.2610 Acc: 0.1583
Epoch 4/10
train | Loss: 2.1890 Acc: 0.1333
test | Loss: 2.5736 Acc: 0.1333
Epoch 5/10
train | Loss: 2.1477 Acc: 0.1677
test | Loss: 4.2125 Acc: 0.2000
Epoch 6/10
train | Loss: 2.0829 Acc: 0.2000
test | Loss: 3.3767 Acc: 0.1250
Epoch 7/10
train | Loss: 2.0939 Acc: 0.1917
test | Loss: 5.2311 Acc: 0.2250
Epoch 8/10
train | Loss: 1.9979 Acc: 0.2146
test | Loss: 7.3176 Acc: 0.3000
Epoch 9/10
train | Loss: 1.9696 Acc: 0.2604
test | Loss: 8.0537 Acc: 0.2750
Epoch 10/10
train | Loss: 1.9693 Acc: 0.2156
test | Loss: 9.5507 Acc: 0.2833
<__main__.TrainModel at 0x7f13af87f550>

test | Loss: 3.5090 Acc: 0.3250

ResNet152

Epoch 1/10
train | Loss: 3.2848 Acc: 0.1240
test | Loss: 3.0465 Acc: 0.1333
Epoch 2/10
train | Loss: 2.3232 Acc: 0.1427
test | Loss: 15.5570 Acc: 0.1500
Epoch 3/10
train | Loss: 2.2480 Acc: 0.1323
test | Loss: 12.6604 Acc: 0.1167
Epoch 4/10
train | Loss: 2.2143 Acc: 0.1573
test | Loss: 3.0900 Acc: 0.1750
Epoch 5/10
train | Loss: 2.1862 Acc: 0.1625
test | Loss: 2.9063 Acc: 0.2167
Epoch 6/10
train | Loss: 2.0490 Acc: 0.1979
test | Loss: 6.5732 Acc: 0.2250
Epoch 7/10
train | Loss: 2.0593 Acc: 0.1708
test | Loss: 5.1629 Acc: 0.2500
Epoch 8/10
train | Loss: 2.0595 Acc: 0.1833
test | Loss: 3.1081 Acc: 0.2917
Epoch 9/10
train | Loss: 2.0187 Acc: 0.2115
test | Loss: 4.8565 Acc: 0.3250
Epoch 10/10
train | Loss: 2.0304 Acc: 0.2135
test | Loss: 6.7477 Acc: 0.1750
<__main__.TrainModel at 0x7f13af7ff550>

test | Loss: 4.1544 Acc: 0.2917

---

결론

층이 깊어질수록 성능이 안 좋아지는 것을 볼 수 있다. 이유는 깊은 층인만큼 epoch가 더 높아야하는데 그렇지 않기 때문이라고 생각한다. 그리고 데이터도 6동작에 총 1200장으로 적기 때문에 이 요인도 영향을 끼쳤을 거라고 추측된다.

최종 모델 성능 검증을 위해 한번씩 더 검토해보겠지만 데이터를 많이 준비해놓지 않는다면 ResNet18을 사용할 것 같다.