early stopping

출처: gpt

 

좋아, 딥러닝에서 자주 쓰는 **early stopping** 개념을 간단하고 직관적으로 설명해줄게!

---

### 🛑 Early Stopping이란?

> **모델이 더 이상 개선되지 않으면 학습을 조기 종료하는 기법**이야.

---

### 🤔 왜 쓰는 걸까?
- **오버피팅(overfitting)** 방지:  
  너무 오래 학습하면 모델이 학습 데이터에 과하게 맞춰져서,  
  validation 성능은 오히려 **나빠질 수 있어**.

- **학습 시간 절약**:  
  이미 최적 성능을 낸 뒤에는 **더 돌려봐야 의미 없음** → 시간 낭비 방지.

---

### 💡 어떻게 동작하냐면?
보통은 아래 로직을 써:

1. 매 epoch마다 **validation loss (또는 accuracy)** 를 확인해.
2. `patience` 만큼의 epoch 동안 성능이 더 이상 좋아지지 않으면 **학습 중단**!

예시:

```python
patience = 3
best_loss = float('inf')
counter = 0

for epoch in range(num_epochs):
    train()  # 학습
    val_loss = validate()

    if val_loss < best_loss:
        best_loss = val_loss
        counter = 0  # 성능이 좋아지면 초기화
        save_model()
    else:
        counter += 1
        if counter >= patience:
            print("Early stopping!")
            break
```

---

### 🔧 보통 이런 걸 설정해:
| 파라미터 | 설명 |
|----------|------|
| `patience` | 몇 번까지 성능 안 좋아져도 참을지 |
| `monitor` | 기준 지표 (e.g., val loss, val accuracy) |
| `mode` | `min` (loss일 때) 또는 `max` (accuracy일 때) |

---

### 📈 시각적으로 보면:
초반에는 validation loss가 계속 감소하다가,  
어느 순간부터는 더 이상 개선되지 않고 다시 올라가 → 그 지점에서 멈추는 게 early stopping!

---

필요하면 PyTorch 코드에 적용하는 방법도 알려줄 수 있어! 적용해보고 싶어?