Modeling - [Try] ResNet Image_size Adjustment

[UGeunJi]

기존 훈련 데이터들의 이미지 사이즈는 224X224였다. 여기서 2배인 448X448로 resize해서 훈련시켜보았다.

Model: ResNet18, epoch=20

Epoch 1/20
train | Loss: 2.2828 Acc: 0.3323
test | Loss: 2.0976 Acc: 0.2583
Epoch 2/20
train | Loss: 1.6557 Acc: 0.4490
test | Loss: 6.1053 Acc: 0.3417
Epoch 3/20
train | Loss: 1.1846 Acc: 0.6198
test | Loss: 0.6579 Acc: 0.7417
Epoch 4/20
train | Loss: 0.5334 Acc: 0.8177
test | Loss: 0.7086 Acc: 0.7333
Epoch 5/20
train | Loss: 0.2604 Acc: 0.9094
test | Loss: 0.1945 Acc: 0.9167
Epoch 6/20
train | Loss: 0.1374 Acc: 0.9531
test | Loss: 1.0281 Acc: 0.6500
Epoch 7/20
train | Loss: 0.2831 Acc: 0.9156
test | Loss: 0.1254 Acc: 0.9500
Epoch 8/20
train | Loss: 0.1253 Acc: 0.9615
test | Loss: 0.3293 Acc: 0.9167
Epoch 9/20
train | Loss: 0.1768 Acc: 0.9510
test | Loss: 0.1805 Acc: 0.9417
Epoch 10/20
train | Loss: 0.1093 Acc: 0.9667
test | Loss: 0.2632 Acc: 0.9000
Epoch 11/20
train | Loss: 0.0364 Acc: 0.9917
test | Loss: 0.1131 Acc: 0.9750
Epoch 12/20
train | Loss: 0.0282 Acc: 0.9948
test | Loss: 0.0403 Acc: 0.9917
Epoch 13/20
train | Loss: 0.0158 Acc: 0.9958
test | Loss: 0.2186 Acc: 0.9417
Epoch 14/20
train | Loss: 0.0134 Acc: 0.9938
test | Loss: 0.0342 Acc: 0.9833
Epoch 15/20
train | Loss: 0.0282 Acc: 0.9948
test | Loss: 0.0381 Acc: 0.9833
Epoch 16/20
train | Loss: 0.0051 Acc: 0.9979
test | Loss: 0.1442 Acc: 0.9750
Epoch 17/20
train | Loss: 0.0308 Acc: 0.9948
test | Loss: 0.0528 Acc: 0.9750
Epoch 18/20
train | Loss: 0.0430 Acc: 0.9896
test | Loss: 0.0813 Acc: 0.9750
Epoch 19/20
train | Loss: 0.0058 Acc: 1.0000
test | Loss: 0.0688 Acc: 0.9667
Epoch 20/20
train | Loss: 0.0024 Acc: 1.0000
test | Loss: 0.0359 Acc: 0.9833
<__main__.TrainModel at 0x7f7d25d043a0>

test | Loss: 0.1191 Acc: 0.9833

성능과 오차가 굉장히 높게 나왔다. 훈련데이터의 정확률을 1로 모두 맞췄다는 것이다. 수상하지만 성능이 뛰어난 것을 확인할 수 있었다.