Early Stopping을 시키는 데에 있어서 이해가 가지 않는 점이 있어서 질문을 남기게 되었습니다.

아래와 같은 Callback을 선언해서 compile에 넣어주고, 학습을 진행했습니다.

early_stopping_cb = keras.callbacks.EarlyStopping(patience=6, verbose=1,
                                                  restore_best_weights=True)

monitoring과 mode, min_delta를 따로 지정하지 않았으니 val_loss를 모니터링하며 mode는 auto로 loss이기 때문에 자동으로 min으로 들어갑니다. min_delta는 따로 지정하지 않아 0입니다.

이와 관련 keras 구현 github 1947번째 줄 (https://github.com/keras-team/keras/blob/master/keras/callbacks.py#L2120)

그리고, patience에 대한 설명도 github 주석 부분을 보면 consecutive(연속적인) 에포크에서 개선이 없으면 멈춘다고 되어있습니다. 하지만, 아래 에포크에서는 연속적으로 6번 감소가 되지 않는 부분이 발견되지 않아서 왜 멈추는 것인지 이해가 되지 않아 글을 남깁니다. Early Stopping이 되는 데에 다른 이유가 있는 것인지 궁금합니다.

Epoch 1/50
1882/1882 [==============================] - 880s 467ms/step - loss: 1.2263 - accuracy: 0.6585 - val_loss: 0.9816 - val_accuracy: 0.7237 - lr: 0.0010
Epoch 2/50
1882/1882 [==============================] - 874s 465ms/step - loss: 1.0468 - accuracy: 0.7031 - val_loss: 1.0004 - val_accuracy: 0.7237 - lr: 0.0010
Epoch 3/50
1882/1882 [==============================] - 871s 463ms/step - loss: 0.9898 - accuracy: 0.7207 - val_loss: 1.0415 - val_accuracy: 0.6707 - lr: 0.0010
Epoch 4/50
1882/1882 [==============================] - 877s 466ms/step - loss: 0.8964 - accuracy: 0.7392 - val_loss: 0.8319 - val_accuracy: 0.7752 - lr: 0.0010
Epoch 5/50
1882/1882 [==============================] - 881s 468ms/step - loss: 0.8093 - accuracy: 0.7508 - val_loss: 0.8079 - val_accuracy: 0.7399 - lr: 0.0010
Epoch 6/50
1882/1882 [==============================] - 867s 461ms/step - loss: 0.7608 - accuracy: 0.7607 - val_loss: 1.1483 - val_accuracy: 0.5458 - lr: 0.0010
Epoch 7/50
1882/1882 [==============================] - 873s 464ms/step - loss: 0.7202 - accuracy: 0.7666 - val_loss: 1.3388 - val_accuracy: 0.4969 - lr: 0.0010
Epoch 8/50
1882/1882 [==============================] - ETA: 0s - loss: 0.6887 - accuracy: 0.7738
Epoch 8: ReduceLROnPlateau reducing learning rate to 0.0005000000237487257.
1882/1882 [==============================] - 870s 462ms/step - loss: 0.6887 - accuracy: 0.7738 - val_loss: 0.7496 - val_accuracy: 0.7268 - lr: 0.0010
Epoch 9/50
1882/1882 [==============================] - 850s 452ms/step - loss: 0.6010 - accuracy: 0.7933 - val_loss: 0.7776 - val_accuracy: 0.6899 - lr: 5.0000e-04
Epoch 10/50
1882/1882 [==============================] - 853s 453ms/step - loss: 0.5711 - accuracy: 0.8047 - val_loss: 0.5706 - val_accuracy: 0.7950 - lr: 5.0000e-04
Epoch 11/50
1882/1882 [==============================] - 870s 462ms/step - loss: 0.5723 - accuracy: 0.8082 - val_loss: 0.5402 - val_accuracy: 0.8257 - lr: 5.0000e-04
Epoch 12/50
1882/1882 [==============================] - 857s 455ms/step - loss: 0.5641 - accuracy: 0.8085 - val_loss: 1.2622 - val_accuracy: 0.5801 - lr: 5.0000e-04
Epoch 13/50
1882/1882 [==============================] - 853s 453ms/step - loss: 0.5574 - accuracy: 0.8182 - val_loss: 0.7001 - val_accuracy: 0.7253 - lr: 5.0000e-04
Epoch 14/50
1882/1882 [==============================] - 862s 458ms/step - loss: 0.5521 - accuracy: 0.8200 - val_loss: 0.6680 - val_accuracy: 0.7419 - lr: 5.0000e-04
Epoch 15/50
1882/1882 [==============================] - ETA: 0s - loss: 0.5398 - accuracy: 0.8269
Epoch 15: ReduceLROnPlateau reducing learning rate to 0.0002500000118743628.
1882/1882 [==============================] - 840s 446ms/step - loss: 0.5398 - accuracy: 0.8269 - val_loss: 0.6545 - val_accuracy: 0.7716 - lr: 5.0000e-04
Epoch 16/50
1882/1882 [==============================] - 845s 449ms/step - loss: 0.4866 - accuracy: 0.8461 - val_loss: 0.4931 - val_accuracy: 0.8278 - lr: 2.5000e-04
Epoch 17/50
1882/1882 [==============================] - 811s 431ms/step - loss: 0.4551 - accuracy: 0.8570 - val_loss: 0.6126 - val_accuracy: 0.7825 - lr: 2.5000e-04
Epoch 18/50
1882/1882 [==============================] - 821s 436ms/step - loss: 0.4466 - accuracy: 0.8624 - val_loss: 0.5223 - val_accuracy: 0.8096 - lr: 2.5000e-04
Epoch 19/50
1882/1882 [==============================] - 816s 434ms/step - loss: 0.4344 - accuracy: 0.8661 - val_loss: 0.4747 - val_accuracy: 0.8403 - lr: 2.5000e-04
Epoch 20/50
1882/1882 [==============================] - 818s 435ms/step - loss: 0.4307 - accuracy: 0.8688 - val_loss: 0.6935 - val_accuracy: 0.7513 - lr: 2.5000e-04
Epoch 21/50
1882/1882 [==============================] - 818s 434ms/step - loss: 0.4252 - accuracy: 0.8751 - val_loss: 0.5897 - val_accuracy: 0.7924 - lr: 2.5000e-04
Epoch 22/50
1882/1882 [==============================] - 826s 439ms/step - loss: 0.4207 - accuracy: 0.8767 - val_loss: 0.4825 - val_accuracy: 0.8439 - lr: 2.5000e-04
Epoch 23/50
1882/1882 [==============================] - 822s 437ms/step - loss: 0.4204 - accuracy: 0.8779 - val_loss: 0.7381 - val_accuracy: 0.7393 - lr: 2.5000e-04
Epoch 24/50
1882/1882 [==============================] - 824s 438ms/step - loss: 0.4115 - accuracy: 0.8836 - val_loss: 0.6477 - val_accuracy: 0.7784 - lr: 2.5000e-04
Epoch 25/50
1882/1882 [==============================] - ETA: 0s - loss: 0.4154 - accuracy: 0.8838Restoring model weights from the end of the best epoch: 19.
1882/1882 [==============================] - 828s 440ms/step - loss: 0.4154 - accuracy: 0.8838 - val_loss: 0.4904 - val_accuracy: 0.8444 - lr: 2.5000e-04
Epoch 25: early stopping